; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5

; FIXME: Currently block machineinstr verifier due to SI BUNDLE pass break physical register liveness. Should remove when the issue is fixed up

; RUN: llc -mtriple=amdgcn -mcpu=tahiti -verify-machineinstrs=0 < %s | FileCheck -check-prefix=SI %s
; RUN: llc -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs=0 < %s | FileCheck -check-prefix=VI %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs=0 < %s | FileCheck -check-prefix=GFX9 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs=0 < %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs=0 < %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s

define <32 x float> @bitcast_v32i32_to_v32f32(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v32f32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB0_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:  .LBB0_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v32f32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB0_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:  .LBB0_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32i32_to_v32f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB0_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11
; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10
; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9
; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8
; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7
; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6
; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5
; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4
; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3
; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:  .LBB0_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32i32_to_v32f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB0_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0
; GFX11-NEXT:  .LBB0_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define inreg <32 x float> @bitcast_v32i32_to_v32f32_scalar(<32 x i32> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32i32_to_v32f32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s10, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s12, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s14, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s42, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s47, v0
; SI-NEXT:    s_cbranch_scc0 .LBB1_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB1_3
; SI-NEXT:  .LBB1_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s6, s6, 3
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s10, s10, 3
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_add_i32 s40, s40, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s42, s42, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s44, s44, 3
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_add_i32 s47, s47, 3
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:  .LBB1_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s47
; SI-NEXT:    v_mov_b32_e32 v15, s46
; SI-NEXT:    v_mov_b32_e32 v16, s45
; SI-NEXT:    v_mov_b32_e32 v17, s44
; SI-NEXT:    v_mov_b32_e32 v18, s43
; SI-NEXT:    v_mov_b32_e32 v19, s42
; SI-NEXT:    v_mov_b32_e32 v20, s41
; SI-NEXT:    v_mov_b32_e32 v21, s40
; SI-NEXT:    v_mov_b32_e32 v22, s15
; SI-NEXT:    v_mov_b32_e32 v23, s14
; SI-NEXT:    v_mov_b32_e32 v24, s13
; SI-NEXT:    v_mov_b32_e32 v25, s12
; SI-NEXT:    v_mov_b32_e32 v26, s11
; SI-NEXT:    v_mov_b32_e32 v27, s10
; SI-NEXT:    v_mov_b32_e32 v28, s9
; SI-NEXT:    v_mov_b32_e32 v29, s8
; SI-NEXT:    v_mov_b32_e32 v30, s7
; SI-NEXT:    v_mov_b32_e32 v31, s6
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB1_4:
; SI-NEXT:    s_branch .LBB1_2
;
; VI-LABEL: bitcast_v32i32_to_v32f32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB1_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB1_3
; VI-NEXT:  .LBB1_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s47, s47, 3
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:  .LBB1_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB1_4:
; VI-NEXT:    s_branch .LBB1_2
;
; GFX9-LABEL: bitcast_v32i32_to_v32f32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB1_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB1_3
; GFX9-NEXT:  .LBB1_2: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    s_add_i32 s9, s9, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    s_add_i32 s13, s13, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s15, s15, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s41, s41, 3
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    s_add_i32 s45, s45, 3
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s47, s47, 3
; GFX9-NEXT:    s_add_i32 s29, s29, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s27, s27, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s25, s25, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s23, s23, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s21, s21, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s19, s19, 3
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    s_add_i32 s17, s17, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:  .LBB1_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB1_4:
; GFX9-NEXT:    s_branch .LBB1_2
;
; GFX11-LABEL: bitcast_v32i32_to_v32f32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB1_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB1_3
; GFX11-NEXT:  .LBB1_2: ; %cmp.true
; GFX11-NEXT:    s_add_i32 s4, s4, 3
; GFX11-NEXT:    s_add_i32 s5, s5, 3
; GFX11-NEXT:    s_add_i32 s6, s6, 3
; GFX11-NEXT:    s_add_i32 s7, s7, 3
; GFX11-NEXT:    s_add_i32 s8, s8, 3
; GFX11-NEXT:    s_add_i32 s9, s9, 3
; GFX11-NEXT:    s_add_i32 s10, s10, 3
; GFX11-NEXT:    s_add_i32 s11, s11, 3
; GFX11-NEXT:    s_add_i32 s12, s12, 3
; GFX11-NEXT:    s_add_i32 s13, s13, 3
; GFX11-NEXT:    s_add_i32 s14, s14, 3
; GFX11-NEXT:    s_add_i32 s15, s15, 3
; GFX11-NEXT:    s_add_i32 s40, s40, 3
; GFX11-NEXT:    s_add_i32 s41, s41, 3
; GFX11-NEXT:    s_add_i32 s29, s29, 3
; GFX11-NEXT:    s_add_i32 s28, s28, 3
; GFX11-NEXT:    s_add_i32 s27, s27, 3
; GFX11-NEXT:    s_add_i32 s26, s26, 3
; GFX11-NEXT:    s_add_i32 s25, s25, 3
; GFX11-NEXT:    s_add_i32 s24, s24, 3
; GFX11-NEXT:    s_add_i32 s23, s23, 3
; GFX11-NEXT:    s_add_i32 s22, s22, 3
; GFX11-NEXT:    s_add_i32 s21, s21, 3
; GFX11-NEXT:    s_add_i32 s20, s20, 3
; GFX11-NEXT:    s_add_i32 s19, s19, 3
; GFX11-NEXT:    s_add_i32 s18, s18, 3
; GFX11-NEXT:    s_add_i32 s17, s17, 3
; GFX11-NEXT:    s_add_i32 s16, s16, 3
; GFX11-NEXT:    s_add_i32 s3, s3, 3
; GFX11-NEXT:    s_add_i32 s2, s2, 3
; GFX11-NEXT:    s_add_i32 s1, s1, 3
; GFX11-NEXT:    s_add_i32 s0, s0, 3
; GFX11-NEXT:  .LBB1_3: ; %end
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB1_4:
; GFX11-NEXT:    s_branch .LBB1_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define <32 x i32> @bitcast_v32f32_to_v32i32(<32 x float> %a, i32 %b) {
; SI-LABEL: bitcast_v32f32_to_v32i32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB2_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; SI-NEXT:  .LBB2_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v32i32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB2_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; VI-NEXT:  .LBB2_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v32i32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB2_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30
; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29
; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28
; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27
; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26
; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25
; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24
; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23
; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22
; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21
; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20
; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19
; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18
; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17
; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16
; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15
; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14
; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13
; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12
; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11
; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10
; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9
; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8
; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7
; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6
; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5
; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4
; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3
; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2
; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0
; GFX9-NEXT:  .LBB2_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v32i32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB2_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
; GFX11-NEXT:  .LBB2_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define inreg <32 x i32> @bitcast_v32f32_to_v32i32_scalar(<32 x float> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32f32_to_v32i32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s36, 0
; SI-NEXT:    v_writelane_b32 v32, s37, 1
; SI-NEXT:    v_writelane_b32 v32, s38, 2
; SI-NEXT:    v_writelane_b32 v32, s39, 3
; SI-NEXT:    v_writelane_b32 v32, s48, 4
; SI-NEXT:    v_writelane_b32 v32, s49, 5
; SI-NEXT:    v_writelane_b32 v32, s50, 6
; SI-NEXT:    v_writelane_b32 v32, s51, 7
; SI-NEXT:    v_writelane_b32 v32, s52, 8
; SI-NEXT:    v_writelane_b32 v32, s53, 9
; SI-NEXT:    v_writelane_b32 v32, s54, 10
; SI-NEXT:    v_writelane_b32 v32, s55, 11
; SI-NEXT:    v_writelane_b32 v32, s64, 12
; SI-NEXT:    v_writelane_b32 v32, s65, 13
; SI-NEXT:    v_writelane_b32 v32, s66, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v32, s67, 15
; SI-NEXT:    s_mov_b32 s49, s29
; SI-NEXT:    s_mov_b32 s48, s28
; SI-NEXT:    s_mov_b32 s47, s27
; SI-NEXT:    s_mov_b32 s46, s26
; SI-NEXT:    s_mov_b32 s45, s25
; SI-NEXT:    s_mov_b32 s44, s24
; SI-NEXT:    s_mov_b32 s43, s23
; SI-NEXT:    s_mov_b32 s42, s22
; SI-NEXT:    s_mov_b32 s41, s21
; SI-NEXT:    s_mov_b32 s40, s20
; SI-NEXT:    s_mov_b32 s39, s19
; SI-NEXT:    s_mov_b32 s38, s18
; SI-NEXT:    s_mov_b32 s37, s17
; SI-NEXT:    s_mov_b32 s36, s16
; SI-NEXT:    v_readfirstlane_b32 s67, v17
; SI-NEXT:    v_readfirstlane_b32 s66, v16
; SI-NEXT:    v_readfirstlane_b32 s65, v15
; SI-NEXT:    v_readfirstlane_b32 s64, v14
; SI-NEXT:    v_readfirstlane_b32 s63, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v12
; SI-NEXT:    v_readfirstlane_b32 s61, v11
; SI-NEXT:    v_readfirstlane_b32 s60, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s58, v8
; SI-NEXT:    v_readfirstlane_b32 s57, v7
; SI-NEXT:    v_readfirstlane_b32 s56, v6
; SI-NEXT:    v_readfirstlane_b32 s55, v5
; SI-NEXT:    v_readfirstlane_b32 s54, v4
; SI-NEXT:    v_readfirstlane_b32 s53, v3
; SI-NEXT:    v_readfirstlane_b32 s52, v2
; SI-NEXT:    v_readfirstlane_b32 s51, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s50, v0
; SI-NEXT:    s_cbranch_scc0 .LBB3_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB3_4
; SI-NEXT:  .LBB3_2: ; %cmp.true
; SI-NEXT:    v_add_f32_e64 v31, s67, 1.0
; SI-NEXT:    v_add_f32_e64 v30, s66, 1.0
; SI-NEXT:    v_add_f32_e64 v29, s65, 1.0
; SI-NEXT:    v_add_f32_e64 v28, s64, 1.0
; SI-NEXT:    v_add_f32_e64 v27, s63, 1.0
; SI-NEXT:    v_add_f32_e64 v26, s62, 1.0
; SI-NEXT:    v_add_f32_e64 v25, s61, 1.0
; SI-NEXT:    v_add_f32_e64 v24, s60, 1.0
; SI-NEXT:    v_add_f32_e64 v23, s59, 1.0
; SI-NEXT:    v_add_f32_e64 v22, s58, 1.0
; SI-NEXT:    v_add_f32_e64 v21, s57, 1.0
; SI-NEXT:    v_add_f32_e64 v20, s56, 1.0
; SI-NEXT:    v_add_f32_e64 v19, s55, 1.0
; SI-NEXT:    v_add_f32_e64 v18, s54, 1.0
; SI-NEXT:    v_add_f32_e64 v17, s53, 1.0
; SI-NEXT:    v_add_f32_e64 v16, s52, 1.0
; SI-NEXT:    v_add_f32_e64 v15, s51, 1.0
; SI-NEXT:    v_add_f32_e64 v14, s50, 1.0
; SI-NEXT:    v_add_f32_e64 v13, s49, 1.0
; SI-NEXT:    v_add_f32_e64 v12, s48, 1.0
; SI-NEXT:    v_add_f32_e64 v11, s47, 1.0
; SI-NEXT:    v_add_f32_e64 v10, s46, 1.0
; SI-NEXT:    v_add_f32_e64 v9, s45, 1.0
; SI-NEXT:    v_add_f32_e64 v8, s44, 1.0
; SI-NEXT:    v_add_f32_e64 v7, s43, 1.0
; SI-NEXT:    v_add_f32_e64 v6, s42, 1.0
; SI-NEXT:    v_add_f32_e64 v5, s41, 1.0
; SI-NEXT:    v_add_f32_e64 v4, s40, 1.0
; SI-NEXT:    v_add_f32_e64 v3, s39, 1.0
; SI-NEXT:    v_add_f32_e64 v2, s38, 1.0
; SI-NEXT:    v_add_f32_e64 v1, s37, 1.0
; SI-NEXT:    v_add_f32_e64 v0, s36, 1.0
; SI-NEXT:    s_branch .LBB3_5
; SI-NEXT:  .LBB3_3:
; SI-NEXT:    s_branch .LBB3_2
; SI-NEXT:  .LBB3_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB3_5: ; %end
; SI-NEXT:    v_readlane_b32 s67, v32, 15
; SI-NEXT:    v_readlane_b32 s66, v32, 14
; SI-NEXT:    v_readlane_b32 s65, v32, 13
; SI-NEXT:    v_readlane_b32 s64, v32, 12
; SI-NEXT:    v_readlane_b32 s55, v32, 11
; SI-NEXT:    v_readlane_b32 s54, v32, 10
; SI-NEXT:    v_readlane_b32 s53, v32, 9
; SI-NEXT:    v_readlane_b32 s52, v32, 8
; SI-NEXT:    v_readlane_b32 s51, v32, 7
; SI-NEXT:    v_readlane_b32 s50, v32, 6
; SI-NEXT:    v_readlane_b32 s49, v32, 5
; SI-NEXT:    v_readlane_b32 s48, v32, 4
; SI-NEXT:    v_readlane_b32 s39, v32, 3
; SI-NEXT:    v_readlane_b32 s38, v32, 2
; SI-NEXT:    v_readlane_b32 s37, v32, 1
; SI-NEXT:    v_readlane_b32 s36, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v32i32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s36, 0
; VI-NEXT:    v_writelane_b32 v32, s37, 1
; VI-NEXT:    v_writelane_b32 v32, s38, 2
; VI-NEXT:    v_writelane_b32 v32, s39, 3
; VI-NEXT:    v_writelane_b32 v32, s48, 4
; VI-NEXT:    v_writelane_b32 v32, s49, 5
; VI-NEXT:    v_writelane_b32 v32, s50, 6
; VI-NEXT:    v_writelane_b32 v32, s51, 7
; VI-NEXT:    v_writelane_b32 v32, s52, 8
; VI-NEXT:    v_writelane_b32 v32, s53, 9
; VI-NEXT:    v_writelane_b32 v32, s54, 10
; VI-NEXT:    v_writelane_b32 v32, s55, 11
; VI-NEXT:    v_writelane_b32 v32, s64, 12
; VI-NEXT:    v_writelane_b32 v32, s65, 13
; VI-NEXT:    v_writelane_b32 v32, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB3_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB3_4
; VI-NEXT:  .LBB3_2: ; %cmp.true
; VI-NEXT:    v_add_f32_e64 v31, s67, 1.0
; VI-NEXT:    v_add_f32_e64 v30, s66, 1.0
; VI-NEXT:    v_add_f32_e64 v29, s65, 1.0
; VI-NEXT:    v_add_f32_e64 v28, s64, 1.0
; VI-NEXT:    v_add_f32_e64 v27, s63, 1.0
; VI-NEXT:    v_add_f32_e64 v26, s62, 1.0
; VI-NEXT:    v_add_f32_e64 v25, s61, 1.0
; VI-NEXT:    v_add_f32_e64 v24, s60, 1.0
; VI-NEXT:    v_add_f32_e64 v23, s59, 1.0
; VI-NEXT:    v_add_f32_e64 v22, s58, 1.0
; VI-NEXT:    v_add_f32_e64 v21, s57, 1.0
; VI-NEXT:    v_add_f32_e64 v20, s56, 1.0
; VI-NEXT:    v_add_f32_e64 v19, s55, 1.0
; VI-NEXT:    v_add_f32_e64 v18, s54, 1.0
; VI-NEXT:    v_add_f32_e64 v17, s53, 1.0
; VI-NEXT:    v_add_f32_e64 v16, s52, 1.0
; VI-NEXT:    v_add_f32_e64 v15, s51, 1.0
; VI-NEXT:    v_add_f32_e64 v14, s50, 1.0
; VI-NEXT:    v_add_f32_e64 v13, s49, 1.0
; VI-NEXT:    v_add_f32_e64 v12, s48, 1.0
; VI-NEXT:    v_add_f32_e64 v11, s47, 1.0
; VI-NEXT:    v_add_f32_e64 v10, s46, 1.0
; VI-NEXT:    v_add_f32_e64 v9, s45, 1.0
; VI-NEXT:    v_add_f32_e64 v8, s44, 1.0
; VI-NEXT:    v_add_f32_e64 v7, s43, 1.0
; VI-NEXT:    v_add_f32_e64 v6, s42, 1.0
; VI-NEXT:    v_add_f32_e64 v5, s41, 1.0
; VI-NEXT:    v_add_f32_e64 v4, s40, 1.0
; VI-NEXT:    v_add_f32_e64 v3, s39, 1.0
; VI-NEXT:    v_add_f32_e64 v2, s38, 1.0
; VI-NEXT:    v_add_f32_e64 v1, s37, 1.0
; VI-NEXT:    v_add_f32_e64 v0, s36, 1.0
; VI-NEXT:    s_branch .LBB3_5
; VI-NEXT:  .LBB3_3:
; VI-NEXT:    s_branch .LBB3_2
; VI-NEXT:  .LBB3_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB3_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v32, 15
; VI-NEXT:    v_readlane_b32 s66, v32, 14
; VI-NEXT:    v_readlane_b32 s65, v32, 13
; VI-NEXT:    v_readlane_b32 s64, v32, 12
; VI-NEXT:    v_readlane_b32 s55, v32, 11
; VI-NEXT:    v_readlane_b32 s54, v32, 10
; VI-NEXT:    v_readlane_b32 s53, v32, 9
; VI-NEXT:    v_readlane_b32 s52, v32, 8
; VI-NEXT:    v_readlane_b32 s51, v32, 7
; VI-NEXT:    v_readlane_b32 s50, v32, 6
; VI-NEXT:    v_readlane_b32 s49, v32, 5
; VI-NEXT:    v_readlane_b32 s48, v32, 4
; VI-NEXT:    v_readlane_b32 s39, v32, 3
; VI-NEXT:    v_readlane_b32 s38, v32, 2
; VI-NEXT:    v_readlane_b32 s37, v32, 1
; VI-NEXT:    v_readlane_b32 s36, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v32i32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB3_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB3_4
; GFX9-NEXT:  .LBB3_2: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e64 v31, s67, 1.0
; GFX9-NEXT:    v_add_f32_e64 v30, s66, 1.0
; GFX9-NEXT:    v_add_f32_e64 v29, s65, 1.0
; GFX9-NEXT:    v_add_f32_e64 v28, s64, 1.0
; GFX9-NEXT:    v_add_f32_e64 v27, s63, 1.0
; GFX9-NEXT:    v_add_f32_e64 v26, s62, 1.0
; GFX9-NEXT:    v_add_f32_e64 v25, s61, 1.0
; GFX9-NEXT:    v_add_f32_e64 v24, s60, 1.0
; GFX9-NEXT:    v_add_f32_e64 v23, s59, 1.0
; GFX9-NEXT:    v_add_f32_e64 v22, s58, 1.0
; GFX9-NEXT:    v_add_f32_e64 v21, s57, 1.0
; GFX9-NEXT:    v_add_f32_e64 v20, s56, 1.0
; GFX9-NEXT:    v_add_f32_e64 v19, s55, 1.0
; GFX9-NEXT:    v_add_f32_e64 v18, s54, 1.0
; GFX9-NEXT:    v_add_f32_e64 v17, s53, 1.0
; GFX9-NEXT:    v_add_f32_e64 v16, s52, 1.0
; GFX9-NEXT:    v_add_f32_e64 v15, s51, 1.0
; GFX9-NEXT:    v_add_f32_e64 v14, s50, 1.0
; GFX9-NEXT:    v_add_f32_e64 v13, s49, 1.0
; GFX9-NEXT:    v_add_f32_e64 v12, s48, 1.0
; GFX9-NEXT:    v_add_f32_e64 v11, s47, 1.0
; GFX9-NEXT:    v_add_f32_e64 v10, s46, 1.0
; GFX9-NEXT:    v_add_f32_e64 v9, s45, 1.0
; GFX9-NEXT:    v_add_f32_e64 v8, s44, 1.0
; GFX9-NEXT:    v_add_f32_e64 v7, s43, 1.0
; GFX9-NEXT:    v_add_f32_e64 v6, s42, 1.0
; GFX9-NEXT:    v_add_f32_e64 v5, s41, 1.0
; GFX9-NEXT:    v_add_f32_e64 v4, s40, 1.0
; GFX9-NEXT:    v_add_f32_e64 v3, s39, 1.0
; GFX9-NEXT:    v_add_f32_e64 v2, s38, 1.0
; GFX9-NEXT:    v_add_f32_e64 v1, s37, 1.0
; GFX9-NEXT:    v_add_f32_e64 v0, s36, 1.0
; GFX9-NEXT:    s_branch .LBB3_5
; GFX9-NEXT:  .LBB3_3:
; GFX9-NEXT:    s_branch .LBB3_2
; GFX9-NEXT:  .LBB3_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB3_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v32i32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB3_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB3_4
; GFX11-NEXT:  .LBB3_2: ; %cmp.true
; GFX11-NEXT:    v_add_f32_e64 v31, s67, 1.0
; GFX11-NEXT:    v_add_f32_e64 v30, s66, 1.0
; GFX11-NEXT:    v_add_f32_e64 v29, s65, 1.0
; GFX11-NEXT:    v_add_f32_e64 v28, s64, 1.0
; GFX11-NEXT:    v_add_f32_e64 v27, s63, 1.0
; GFX11-NEXT:    v_add_f32_e64 v26, s62, 1.0
; GFX11-NEXT:    v_add_f32_e64 v25, s61, 1.0
; GFX11-NEXT:    v_add_f32_e64 v24, s60, 1.0
; GFX11-NEXT:    v_add_f32_e64 v23, s59, 1.0
; GFX11-NEXT:    v_add_f32_e64 v22, s58, 1.0
; GFX11-NEXT:    v_add_f32_e64 v21, s57, 1.0
; GFX11-NEXT:    v_add_f32_e64 v20, s56, 1.0
; GFX11-NEXT:    v_add_f32_e64 v19, s55, 1.0
; GFX11-NEXT:    v_add_f32_e64 v18, s54, 1.0
; GFX11-NEXT:    v_add_f32_e64 v17, s53, 1.0
; GFX11-NEXT:    v_add_f32_e64 v16, s52, 1.0
; GFX11-NEXT:    v_add_f32_e64 v15, s51, 1.0
; GFX11-NEXT:    v_add_f32_e64 v14, s50, 1.0
; GFX11-NEXT:    v_add_f32_e64 v13, s49, 1.0
; GFX11-NEXT:    v_add_f32_e64 v12, s48, 1.0
; GFX11-NEXT:    v_add_f32_e64 v11, s47, 1.0
; GFX11-NEXT:    v_add_f32_e64 v10, s46, 1.0
; GFX11-NEXT:    v_add_f32_e64 v9, s45, 1.0
; GFX11-NEXT:    v_add_f32_e64 v8, s44, 1.0
; GFX11-NEXT:    v_add_f32_e64 v7, s43, 1.0
; GFX11-NEXT:    v_add_f32_e64 v6, s42, 1.0
; GFX11-NEXT:    v_add_f32_e64 v5, s41, 1.0
; GFX11-NEXT:    v_add_f32_e64 v4, s40, 1.0
; GFX11-NEXT:    v_add_f32_e64 v3, s39, 1.0
; GFX11-NEXT:    v_add_f32_e64 v2, s38, 1.0
; GFX11-NEXT:    v_add_f32_e64 v1, s37, 1.0
; GFX11-NEXT:    v_add_f32_e64 v0, s36, 1.0
; GFX11-NEXT:    s_branch .LBB3_5
; GFX11-NEXT:  .LBB3_3:
; GFX11-NEXT:    s_branch .LBB3_2
; GFX11-NEXT:  .LBB3_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB3_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define <16 x i64> @bitcast_v32i32_to_v16i64(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v16i64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB4_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:  .LBB4_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v16i64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB4_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:  .LBB4_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32i32_to_v16i64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB4_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11
; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10
; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9
; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8
; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7
; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6
; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5
; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4
; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3
; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:  .LBB4_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32i32_to_v16i64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB4_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0
; GFX11-NEXT:  .LBB4_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define inreg <16 x i64> @bitcast_v32i32_to_v16i64_scalar(<32 x i32> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32i32_to_v16i64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s10, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s12, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s14, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s42, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s47, v0
; SI-NEXT:    s_cbranch_scc0 .LBB5_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB5_3
; SI-NEXT:  .LBB5_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s6, s6, 3
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s10, s10, 3
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_add_i32 s40, s40, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s42, s42, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s44, s44, 3
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_add_i32 s47, s47, 3
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:  .LBB5_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s47
; SI-NEXT:    v_mov_b32_e32 v15, s46
; SI-NEXT:    v_mov_b32_e32 v16, s45
; SI-NEXT:    v_mov_b32_e32 v17, s44
; SI-NEXT:    v_mov_b32_e32 v18, s43
; SI-NEXT:    v_mov_b32_e32 v19, s42
; SI-NEXT:    v_mov_b32_e32 v20, s41
; SI-NEXT:    v_mov_b32_e32 v21, s40
; SI-NEXT:    v_mov_b32_e32 v22, s15
; SI-NEXT:    v_mov_b32_e32 v23, s14
; SI-NEXT:    v_mov_b32_e32 v24, s13
; SI-NEXT:    v_mov_b32_e32 v25, s12
; SI-NEXT:    v_mov_b32_e32 v26, s11
; SI-NEXT:    v_mov_b32_e32 v27, s10
; SI-NEXT:    v_mov_b32_e32 v28, s9
; SI-NEXT:    v_mov_b32_e32 v29, s8
; SI-NEXT:    v_mov_b32_e32 v30, s7
; SI-NEXT:    v_mov_b32_e32 v31, s6
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB5_4:
; SI-NEXT:    s_branch .LBB5_2
;
; VI-LABEL: bitcast_v32i32_to_v16i64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB5_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB5_3
; VI-NEXT:  .LBB5_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s47, s47, 3
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:  .LBB5_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB5_4:
; VI-NEXT:    s_branch .LBB5_2
;
; GFX9-LABEL: bitcast_v32i32_to_v16i64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB5_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB5_3
; GFX9-NEXT:  .LBB5_2: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    s_add_i32 s9, s9, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    s_add_i32 s13, s13, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s15, s15, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s41, s41, 3
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    s_add_i32 s45, s45, 3
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s47, s47, 3
; GFX9-NEXT:    s_add_i32 s29, s29, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s27, s27, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s25, s25, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s23, s23, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s21, s21, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s19, s19, 3
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    s_add_i32 s17, s17, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:  .LBB5_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB5_4:
; GFX9-NEXT:    s_branch .LBB5_2
;
; GFX11-LABEL: bitcast_v32i32_to_v16i64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB5_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB5_3
; GFX11-NEXT:  .LBB5_2: ; %cmp.true
; GFX11-NEXT:    s_add_i32 s4, s4, 3
; GFX11-NEXT:    s_add_i32 s5, s5, 3
; GFX11-NEXT:    s_add_i32 s6, s6, 3
; GFX11-NEXT:    s_add_i32 s7, s7, 3
; GFX11-NEXT:    s_add_i32 s8, s8, 3
; GFX11-NEXT:    s_add_i32 s9, s9, 3
; GFX11-NEXT:    s_add_i32 s10, s10, 3
; GFX11-NEXT:    s_add_i32 s11, s11, 3
; GFX11-NEXT:    s_add_i32 s12, s12, 3
; GFX11-NEXT:    s_add_i32 s13, s13, 3
; GFX11-NEXT:    s_add_i32 s14, s14, 3
; GFX11-NEXT:    s_add_i32 s15, s15, 3
; GFX11-NEXT:    s_add_i32 s40, s40, 3
; GFX11-NEXT:    s_add_i32 s41, s41, 3
; GFX11-NEXT:    s_add_i32 s29, s29, 3
; GFX11-NEXT:    s_add_i32 s28, s28, 3
; GFX11-NEXT:    s_add_i32 s27, s27, 3
; GFX11-NEXT:    s_add_i32 s26, s26, 3
; GFX11-NEXT:    s_add_i32 s25, s25, 3
; GFX11-NEXT:    s_add_i32 s24, s24, 3
; GFX11-NEXT:    s_add_i32 s23, s23, 3
; GFX11-NEXT:    s_add_i32 s22, s22, 3
; GFX11-NEXT:    s_add_i32 s21, s21, 3
; GFX11-NEXT:    s_add_i32 s20, s20, 3
; GFX11-NEXT:    s_add_i32 s19, s19, 3
; GFX11-NEXT:    s_add_i32 s18, s18, 3
; GFX11-NEXT:    s_add_i32 s17, s17, 3
; GFX11-NEXT:    s_add_i32 s16, s16, 3
; GFX11-NEXT:    s_add_i32 s3, s3, 3
; GFX11-NEXT:    s_add_i32 s2, s2, 3
; GFX11-NEXT:    s_add_i32 s1, s1, 3
; GFX11-NEXT:    s_add_i32 s0, s0, 3
; GFX11-NEXT:  .LBB5_3: ; %end
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB5_4:
; GFX11-NEXT:    s_branch .LBB5_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define <32 x i32> @bitcast_v16i64_to_v32i32(<16 x i64> %a, i32 %b) {
; SI-LABEL: bitcast_v16i64_to_v32i32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB6_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; SI-NEXT:  .LBB6_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16i64_to_v32i32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB6_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; VI-NEXT:  .LBB6_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16i64_to_v32i32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB6_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v28
; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v26
; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v24
; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v22
; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v20
; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v18, vcc, 3, v18
; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v16
; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v14
; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v12
; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v10
; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v8
; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v6
; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v4
; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v2
; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v0
; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX9-NEXT:  .LBB6_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16i64_to_v32i32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB6_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 3
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
; GFX11-NEXT:  .LBB6_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define inreg <32 x i32> @bitcast_v16i64_to_v32i32_scalar(<16 x i64> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16i64_to_v32i32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s10, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s12, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s14, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s42, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s47, v0
; SI-NEXT:    s_cbranch_scc0 .LBB7_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB7_3
; SI-NEXT:  .LBB7_2: ; %cmp.true
; SI-NEXT:    s_add_u32 s7, s7, 3
; SI-NEXT:    s_addc_u32 s6, s6, 0
; SI-NEXT:    s_add_u32 s9, s9, 3
; SI-NEXT:    s_addc_u32 s8, s8, 0
; SI-NEXT:    s_add_u32 s11, s11, 3
; SI-NEXT:    s_addc_u32 s10, s10, 0
; SI-NEXT:    s_add_u32 s13, s13, 3
; SI-NEXT:    s_addc_u32 s12, s12, 0
; SI-NEXT:    s_add_u32 s15, s15, 3
; SI-NEXT:    s_addc_u32 s14, s14, 0
; SI-NEXT:    s_add_u32 s41, s41, 3
; SI-NEXT:    s_addc_u32 s40, s40, 0
; SI-NEXT:    s_add_u32 s43, s43, 3
; SI-NEXT:    s_addc_u32 s42, s42, 0
; SI-NEXT:    s_add_u32 s45, s45, 3
; SI-NEXT:    s_addc_u32 s44, s44, 0
; SI-NEXT:    s_add_u32 s47, s47, 3
; SI-NEXT:    s_addc_u32 s46, s46, 0
; SI-NEXT:    s_add_u32 s28, s28, 3
; SI-NEXT:    s_addc_u32 s29, s29, 0
; SI-NEXT:    s_add_u32 s26, s26, 3
; SI-NEXT:    s_addc_u32 s27, s27, 0
; SI-NEXT:    s_add_u32 s24, s24, 3
; SI-NEXT:    s_addc_u32 s25, s25, 0
; SI-NEXT:    s_add_u32 s22, s22, 3
; SI-NEXT:    s_addc_u32 s23, s23, 0
; SI-NEXT:    s_add_u32 s20, s20, 3
; SI-NEXT:    s_addc_u32 s21, s21, 0
; SI-NEXT:    s_add_u32 s18, s18, 3
; SI-NEXT:    s_addc_u32 s19, s19, 0
; SI-NEXT:    s_add_u32 s16, s16, 3
; SI-NEXT:    s_addc_u32 s17, s17, 0
; SI-NEXT:  .LBB7_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s47
; SI-NEXT:    v_mov_b32_e32 v15, s46
; SI-NEXT:    v_mov_b32_e32 v16, s45
; SI-NEXT:    v_mov_b32_e32 v17, s44
; SI-NEXT:    v_mov_b32_e32 v18, s43
; SI-NEXT:    v_mov_b32_e32 v19, s42
; SI-NEXT:    v_mov_b32_e32 v20, s41
; SI-NEXT:    v_mov_b32_e32 v21, s40
; SI-NEXT:    v_mov_b32_e32 v22, s15
; SI-NEXT:    v_mov_b32_e32 v23, s14
; SI-NEXT:    v_mov_b32_e32 v24, s13
; SI-NEXT:    v_mov_b32_e32 v25, s12
; SI-NEXT:    v_mov_b32_e32 v26, s11
; SI-NEXT:    v_mov_b32_e32 v27, s10
; SI-NEXT:    v_mov_b32_e32 v28, s9
; SI-NEXT:    v_mov_b32_e32 v29, s8
; SI-NEXT:    v_mov_b32_e32 v30, s7
; SI-NEXT:    v_mov_b32_e32 v31, s6
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB7_4:
; SI-NEXT:    s_branch .LBB7_2
;
; VI-LABEL: bitcast_v16i64_to_v32i32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB7_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB7_3
; VI-NEXT:  .LBB7_2: ; %cmp.true
; VI-NEXT:    s_add_u32 s7, s7, 3
; VI-NEXT:    s_addc_u32 s6, s6, 0
; VI-NEXT:    s_add_u32 s9, s9, 3
; VI-NEXT:    s_addc_u32 s8, s8, 0
; VI-NEXT:    s_add_u32 s11, s11, 3
; VI-NEXT:    s_addc_u32 s10, s10, 0
; VI-NEXT:    s_add_u32 s13, s13, 3
; VI-NEXT:    s_addc_u32 s12, s12, 0
; VI-NEXT:    s_add_u32 s15, s15, 3
; VI-NEXT:    s_addc_u32 s14, s14, 0
; VI-NEXT:    s_add_u32 s41, s41, 3
; VI-NEXT:    s_addc_u32 s40, s40, 0
; VI-NEXT:    s_add_u32 s43, s43, 3
; VI-NEXT:    s_addc_u32 s42, s42, 0
; VI-NEXT:    s_add_u32 s45, s45, 3
; VI-NEXT:    s_addc_u32 s44, s44, 0
; VI-NEXT:    s_add_u32 s47, s47, 3
; VI-NEXT:    s_addc_u32 s46, s46, 0
; VI-NEXT:    s_add_u32 s28, s28, 3
; VI-NEXT:    s_addc_u32 s29, s29, 0
; VI-NEXT:    s_add_u32 s26, s26, 3
; VI-NEXT:    s_addc_u32 s27, s27, 0
; VI-NEXT:    s_add_u32 s24, s24, 3
; VI-NEXT:    s_addc_u32 s25, s25, 0
; VI-NEXT:    s_add_u32 s22, s22, 3
; VI-NEXT:    s_addc_u32 s23, s23, 0
; VI-NEXT:    s_add_u32 s20, s20, 3
; VI-NEXT:    s_addc_u32 s21, s21, 0
; VI-NEXT:    s_add_u32 s18, s18, 3
; VI-NEXT:    s_addc_u32 s19, s19, 0
; VI-NEXT:    s_add_u32 s16, s16, 3
; VI-NEXT:    s_addc_u32 s17, s17, 0
; VI-NEXT:  .LBB7_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB7_4:
; VI-NEXT:    s_branch .LBB7_2
;
; GFX9-LABEL: bitcast_v16i64_to_v32i32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB7_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB7_3
; GFX9-NEXT:  .LBB7_2: ; %cmp.true
; GFX9-NEXT:    s_add_u32 s7, s7, 3
; GFX9-NEXT:    s_addc_u32 s6, s6, 0
; GFX9-NEXT:    s_add_u32 s9, s9, 3
; GFX9-NEXT:    s_addc_u32 s8, s8, 0
; GFX9-NEXT:    s_add_u32 s11, s11, 3
; GFX9-NEXT:    s_addc_u32 s10, s10, 0
; GFX9-NEXT:    s_add_u32 s13, s13, 3
; GFX9-NEXT:    s_addc_u32 s12, s12, 0
; GFX9-NEXT:    s_add_u32 s15, s15, 3
; GFX9-NEXT:    s_addc_u32 s14, s14, 0
; GFX9-NEXT:    s_add_u32 s41, s41, 3
; GFX9-NEXT:    s_addc_u32 s40, s40, 0
; GFX9-NEXT:    s_add_u32 s43, s43, 3
; GFX9-NEXT:    s_addc_u32 s42, s42, 0
; GFX9-NEXT:    s_add_u32 s45, s45, 3
; GFX9-NEXT:    s_addc_u32 s44, s44, 0
; GFX9-NEXT:    s_add_u32 s47, s47, 3
; GFX9-NEXT:    s_addc_u32 s46, s46, 0
; GFX9-NEXT:    s_add_u32 s28, s28, 3
; GFX9-NEXT:    s_addc_u32 s29, s29, 0
; GFX9-NEXT:    s_add_u32 s26, s26, 3
; GFX9-NEXT:    s_addc_u32 s27, s27, 0
; GFX9-NEXT:    s_add_u32 s24, s24, 3
; GFX9-NEXT:    s_addc_u32 s25, s25, 0
; GFX9-NEXT:    s_add_u32 s22, s22, 3
; GFX9-NEXT:    s_addc_u32 s23, s23, 0
; GFX9-NEXT:    s_add_u32 s20, s20, 3
; GFX9-NEXT:    s_addc_u32 s21, s21, 0
; GFX9-NEXT:    s_add_u32 s18, s18, 3
; GFX9-NEXT:    s_addc_u32 s19, s19, 0
; GFX9-NEXT:    s_add_u32 s16, s16, 3
; GFX9-NEXT:    s_addc_u32 s17, s17, 0
; GFX9-NEXT:  .LBB7_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB7_4:
; GFX9-NEXT:    s_branch .LBB7_2
;
; GFX11-LABEL: bitcast_v16i64_to_v32i32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB7_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB7_3
; GFX11-NEXT:  .LBB7_2: ; %cmp.true
; GFX11-NEXT:    s_add_u32 s5, s5, 3
; GFX11-NEXT:    s_addc_u32 s4, s4, 0
; GFX11-NEXT:    s_add_u32 s7, s7, 3
; GFX11-NEXT:    s_addc_u32 s6, s6, 0
; GFX11-NEXT:    s_add_u32 s9, s9, 3
; GFX11-NEXT:    s_addc_u32 s8, s8, 0
; GFX11-NEXT:    s_add_u32 s11, s11, 3
; GFX11-NEXT:    s_addc_u32 s10, s10, 0
; GFX11-NEXT:    s_add_u32 s13, s13, 3
; GFX11-NEXT:    s_addc_u32 s12, s12, 0
; GFX11-NEXT:    s_add_u32 s15, s15, 3
; GFX11-NEXT:    s_addc_u32 s14, s14, 0
; GFX11-NEXT:    s_add_u32 s41, s41, 3
; GFX11-NEXT:    s_addc_u32 s40, s40, 0
; GFX11-NEXT:    s_add_u32 s28, s28, 3
; GFX11-NEXT:    s_addc_u32 s29, s29, 0
; GFX11-NEXT:    s_add_u32 s26, s26, 3
; GFX11-NEXT:    s_addc_u32 s27, s27, 0
; GFX11-NEXT:    s_add_u32 s24, s24, 3
; GFX11-NEXT:    s_addc_u32 s25, s25, 0
; GFX11-NEXT:    s_add_u32 s22, s22, 3
; GFX11-NEXT:    s_addc_u32 s23, s23, 0
; GFX11-NEXT:    s_add_u32 s20, s20, 3
; GFX11-NEXT:    s_addc_u32 s21, s21, 0
; GFX11-NEXT:    s_add_u32 s18, s18, 3
; GFX11-NEXT:    s_addc_u32 s19, s19, 0
; GFX11-NEXT:    s_add_u32 s16, s16, 3
; GFX11-NEXT:    s_addc_u32 s17, s17, 0
; GFX11-NEXT:    s_add_u32 s2, s2, 3
; GFX11-NEXT:    s_addc_u32 s3, s3, 0
; GFX11-NEXT:    s_add_u32 s0, s0, 3
; GFX11-NEXT:    s_addc_u32 s1, s1, 0
; GFX11-NEXT:  .LBB7_3: ; %end
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB7_4:
; GFX11-NEXT:    s_branch .LBB7_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define <16 x double> @bitcast_v32i32_to_v16f64(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v16f64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB8_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:  .LBB8_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v16f64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB8_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:  .LBB8_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32i32_to_v16f64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB8_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11
; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10
; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9
; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8
; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7
; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6
; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5
; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4
; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3
; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:  .LBB8_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32i32_to_v16f64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB8_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0
; GFX11-NEXT:  .LBB8_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define inreg <16 x double> @bitcast_v32i32_to_v16f64_scalar(<32 x i32> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32i32_to_v16f64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s10, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s12, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s14, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s42, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s47, v0
; SI-NEXT:    s_cbranch_scc0 .LBB9_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB9_3
; SI-NEXT:  .LBB9_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s6, s6, 3
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s10, s10, 3
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_add_i32 s40, s40, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s42, s42, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s44, s44, 3
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_add_i32 s47, s47, 3
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:  .LBB9_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s47
; SI-NEXT:    v_mov_b32_e32 v15, s46
; SI-NEXT:    v_mov_b32_e32 v16, s45
; SI-NEXT:    v_mov_b32_e32 v17, s44
; SI-NEXT:    v_mov_b32_e32 v18, s43
; SI-NEXT:    v_mov_b32_e32 v19, s42
; SI-NEXT:    v_mov_b32_e32 v20, s41
; SI-NEXT:    v_mov_b32_e32 v21, s40
; SI-NEXT:    v_mov_b32_e32 v22, s15
; SI-NEXT:    v_mov_b32_e32 v23, s14
; SI-NEXT:    v_mov_b32_e32 v24, s13
; SI-NEXT:    v_mov_b32_e32 v25, s12
; SI-NEXT:    v_mov_b32_e32 v26, s11
; SI-NEXT:    v_mov_b32_e32 v27, s10
; SI-NEXT:    v_mov_b32_e32 v28, s9
; SI-NEXT:    v_mov_b32_e32 v29, s8
; SI-NEXT:    v_mov_b32_e32 v30, s7
; SI-NEXT:    v_mov_b32_e32 v31, s6
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB9_4:
; SI-NEXT:    s_branch .LBB9_2
;
; VI-LABEL: bitcast_v32i32_to_v16f64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB9_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB9_3
; VI-NEXT:  .LBB9_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s47, s47, 3
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:  .LBB9_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB9_4:
; VI-NEXT:    s_branch .LBB9_2
;
; GFX9-LABEL: bitcast_v32i32_to_v16f64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB9_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB9_3
; GFX9-NEXT:  .LBB9_2: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    s_add_i32 s9, s9, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    s_add_i32 s13, s13, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s15, s15, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s41, s41, 3
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    s_add_i32 s45, s45, 3
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s47, s47, 3
; GFX9-NEXT:    s_add_i32 s29, s29, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s27, s27, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s25, s25, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s23, s23, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s21, s21, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s19, s19, 3
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    s_add_i32 s17, s17, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:  .LBB9_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB9_4:
; GFX9-NEXT:    s_branch .LBB9_2
;
; GFX11-LABEL: bitcast_v32i32_to_v16f64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB9_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB9_3
; GFX11-NEXT:  .LBB9_2: ; %cmp.true
; GFX11-NEXT:    s_add_i32 s4, s4, 3
; GFX11-NEXT:    s_add_i32 s5, s5, 3
; GFX11-NEXT:    s_add_i32 s6, s6, 3
; GFX11-NEXT:    s_add_i32 s7, s7, 3
; GFX11-NEXT:    s_add_i32 s8, s8, 3
; GFX11-NEXT:    s_add_i32 s9, s9, 3
; GFX11-NEXT:    s_add_i32 s10, s10, 3
; GFX11-NEXT:    s_add_i32 s11, s11, 3
; GFX11-NEXT:    s_add_i32 s12, s12, 3
; GFX11-NEXT:    s_add_i32 s13, s13, 3
; GFX11-NEXT:    s_add_i32 s14, s14, 3
; GFX11-NEXT:    s_add_i32 s15, s15, 3
; GFX11-NEXT:    s_add_i32 s40, s40, 3
; GFX11-NEXT:    s_add_i32 s41, s41, 3
; GFX11-NEXT:    s_add_i32 s29, s29, 3
; GFX11-NEXT:    s_add_i32 s28, s28, 3
; GFX11-NEXT:    s_add_i32 s27, s27, 3
; GFX11-NEXT:    s_add_i32 s26, s26, 3
; GFX11-NEXT:    s_add_i32 s25, s25, 3
; GFX11-NEXT:    s_add_i32 s24, s24, 3
; GFX11-NEXT:    s_add_i32 s23, s23, 3
; GFX11-NEXT:    s_add_i32 s22, s22, 3
; GFX11-NEXT:    s_add_i32 s21, s21, 3
; GFX11-NEXT:    s_add_i32 s20, s20, 3
; GFX11-NEXT:    s_add_i32 s19, s19, 3
; GFX11-NEXT:    s_add_i32 s18, s18, 3
; GFX11-NEXT:    s_add_i32 s17, s17, 3
; GFX11-NEXT:    s_add_i32 s16, s16, 3
; GFX11-NEXT:    s_add_i32 s3, s3, 3
; GFX11-NEXT:    s_add_i32 s2, s2, 3
; GFX11-NEXT:    s_add_i32 s1, s1, 3
; GFX11-NEXT:    s_add_i32 s0, s0, 3
; GFX11-NEXT:  .LBB9_3: ; %end
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB9_4:
; GFX11-NEXT:    s_branch .LBB9_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define <32 x i32> @bitcast_v16f64_to_v32i32(<16 x double> %a, i32 %b) {
; SI-LABEL: bitcast_v16f64_to_v32i32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB10_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; SI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; SI-NEXT:  .LBB10_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v32i32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB10_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; VI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; VI-NEXT:  .LBB10_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v32i32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB10_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX9-NEXT:  .LBB10_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v32i32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB10_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX11-NEXT:  .LBB10_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define inreg <32 x i32> @bitcast_v16f64_to_v32i32_scalar(<16 x double> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16f64_to_v32i32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s36, 0
; SI-NEXT:    v_writelane_b32 v32, s37, 1
; SI-NEXT:    v_writelane_b32 v32, s38, 2
; SI-NEXT:    v_writelane_b32 v32, s39, 3
; SI-NEXT:    v_writelane_b32 v32, s48, 4
; SI-NEXT:    v_writelane_b32 v32, s49, 5
; SI-NEXT:    v_writelane_b32 v32, s50, 6
; SI-NEXT:    v_writelane_b32 v32, s51, 7
; SI-NEXT:    v_writelane_b32 v32, s52, 8
; SI-NEXT:    v_writelane_b32 v32, s53, 9
; SI-NEXT:    v_writelane_b32 v32, s54, 10
; SI-NEXT:    v_writelane_b32 v32, s55, 11
; SI-NEXT:    v_writelane_b32 v32, s64, 12
; SI-NEXT:    v_writelane_b32 v32, s65, 13
; SI-NEXT:    v_writelane_b32 v32, s66, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v32, s67, 15
; SI-NEXT:    s_mov_b32 s49, s29
; SI-NEXT:    s_mov_b32 s48, s28
; SI-NEXT:    s_mov_b32 s47, s27
; SI-NEXT:    s_mov_b32 s46, s26
; SI-NEXT:    s_mov_b32 s45, s25
; SI-NEXT:    s_mov_b32 s44, s24
; SI-NEXT:    s_mov_b32 s43, s23
; SI-NEXT:    s_mov_b32 s42, s22
; SI-NEXT:    s_mov_b32 s41, s21
; SI-NEXT:    s_mov_b32 s40, s20
; SI-NEXT:    s_mov_b32 s39, s19
; SI-NEXT:    s_mov_b32 s38, s18
; SI-NEXT:    s_mov_b32 s37, s17
; SI-NEXT:    s_mov_b32 s36, s16
; SI-NEXT:    v_readfirstlane_b32 s67, v17
; SI-NEXT:    v_readfirstlane_b32 s66, v16
; SI-NEXT:    v_readfirstlane_b32 s65, v15
; SI-NEXT:    v_readfirstlane_b32 s64, v14
; SI-NEXT:    v_readfirstlane_b32 s63, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v12
; SI-NEXT:    v_readfirstlane_b32 s61, v11
; SI-NEXT:    v_readfirstlane_b32 s60, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s58, v8
; SI-NEXT:    v_readfirstlane_b32 s57, v7
; SI-NEXT:    v_readfirstlane_b32 s56, v6
; SI-NEXT:    v_readfirstlane_b32 s55, v5
; SI-NEXT:    v_readfirstlane_b32 s54, v4
; SI-NEXT:    v_readfirstlane_b32 s53, v3
; SI-NEXT:    v_readfirstlane_b32 s52, v2
; SI-NEXT:    v_readfirstlane_b32 s51, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s50, v0
; SI-NEXT:    s_cbranch_scc0 .LBB11_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB11_4
; SI-NEXT:  .LBB11_2: ; %cmp.true
; SI-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; SI-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; SI-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; SI-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; SI-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; SI-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; SI-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; SI-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; SI-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; SI-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; SI-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; SI-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; SI-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; SI-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; SI-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; SI-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; SI-NEXT:    s_branch .LBB11_5
; SI-NEXT:  .LBB11_3:
; SI-NEXT:    s_branch .LBB11_2
; SI-NEXT:  .LBB11_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB11_5: ; %end
; SI-NEXT:    v_readlane_b32 s67, v32, 15
; SI-NEXT:    v_readlane_b32 s66, v32, 14
; SI-NEXT:    v_readlane_b32 s65, v32, 13
; SI-NEXT:    v_readlane_b32 s64, v32, 12
; SI-NEXT:    v_readlane_b32 s55, v32, 11
; SI-NEXT:    v_readlane_b32 s54, v32, 10
; SI-NEXT:    v_readlane_b32 s53, v32, 9
; SI-NEXT:    v_readlane_b32 s52, v32, 8
; SI-NEXT:    v_readlane_b32 s51, v32, 7
; SI-NEXT:    v_readlane_b32 s50, v32, 6
; SI-NEXT:    v_readlane_b32 s49, v32, 5
; SI-NEXT:    v_readlane_b32 s48, v32, 4
; SI-NEXT:    v_readlane_b32 s39, v32, 3
; SI-NEXT:    v_readlane_b32 s38, v32, 2
; SI-NEXT:    v_readlane_b32 s37, v32, 1
; SI-NEXT:    v_readlane_b32 s36, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v32i32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s36, 0
; VI-NEXT:    v_writelane_b32 v32, s37, 1
; VI-NEXT:    v_writelane_b32 v32, s38, 2
; VI-NEXT:    v_writelane_b32 v32, s39, 3
; VI-NEXT:    v_writelane_b32 v32, s48, 4
; VI-NEXT:    v_writelane_b32 v32, s49, 5
; VI-NEXT:    v_writelane_b32 v32, s50, 6
; VI-NEXT:    v_writelane_b32 v32, s51, 7
; VI-NEXT:    v_writelane_b32 v32, s52, 8
; VI-NEXT:    v_writelane_b32 v32, s53, 9
; VI-NEXT:    v_writelane_b32 v32, s54, 10
; VI-NEXT:    v_writelane_b32 v32, s55, 11
; VI-NEXT:    v_writelane_b32 v32, s64, 12
; VI-NEXT:    v_writelane_b32 v32, s65, 13
; VI-NEXT:    v_writelane_b32 v32, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB11_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB11_4
; VI-NEXT:  .LBB11_2: ; %cmp.true
; VI-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; VI-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; VI-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; VI-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; VI-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; VI-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; VI-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; VI-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; VI-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; VI-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; VI-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; VI-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; VI-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; VI-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; VI-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; VI-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; VI-NEXT:    s_branch .LBB11_5
; VI-NEXT:  .LBB11_3:
; VI-NEXT:    s_branch .LBB11_2
; VI-NEXT:  .LBB11_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB11_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v32, 15
; VI-NEXT:    v_readlane_b32 s66, v32, 14
; VI-NEXT:    v_readlane_b32 s65, v32, 13
; VI-NEXT:    v_readlane_b32 s64, v32, 12
; VI-NEXT:    v_readlane_b32 s55, v32, 11
; VI-NEXT:    v_readlane_b32 s54, v32, 10
; VI-NEXT:    v_readlane_b32 s53, v32, 9
; VI-NEXT:    v_readlane_b32 s52, v32, 8
; VI-NEXT:    v_readlane_b32 s51, v32, 7
; VI-NEXT:    v_readlane_b32 s50, v32, 6
; VI-NEXT:    v_readlane_b32 s49, v32, 5
; VI-NEXT:    v_readlane_b32 s48, v32, 4
; VI-NEXT:    v_readlane_b32 s39, v32, 3
; VI-NEXT:    v_readlane_b32 s38, v32, 2
; VI-NEXT:    v_readlane_b32 s37, v32, 1
; VI-NEXT:    v_readlane_b32 s36, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v32i32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB11_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB11_4
; GFX9-NEXT:  .LBB11_2: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; GFX9-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; GFX9-NEXT:    s_branch .LBB11_5
; GFX9-NEXT:  .LBB11_3:
; GFX9-NEXT:    s_branch .LBB11_2
; GFX9-NEXT:  .LBB11_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB11_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v32i32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB11_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB11_4
; GFX11-NEXT:  .LBB11_2: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; GFX11-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; GFX11-NEXT:    s_branch .LBB11_5
; GFX11-NEXT:  .LBB11_3:
; GFX11-NEXT:    s_branch .LBB11_2
; GFX11-NEXT:  .LBB11_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB11_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v128i8:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB12_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v53, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v52, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB12_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB12_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 16
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v53, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v52, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB12_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v45
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v1, v1, v45
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v43
; SI-NEXT:    v_lshlrev_b32_e32 v41, 24, v41
; SI-NEXT:    v_or_b32_e32 v41, v41, v43
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v41
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v34
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v33
; SI-NEXT:    v_or_b32_e32 v2, v33, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v55
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v53
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v51
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v61
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v48
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v38
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v37
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v60
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v58
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v7
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v57
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v47
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v9
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v46
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v44
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v42
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v11
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v40
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v52
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v13
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v39
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v15
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v17
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v27
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v28
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v29
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v30
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v31
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v128i8:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB12_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v58, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v56, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v1
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:  .LBB12_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB12_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v32
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v58, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v56, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v1
; VI-NEXT:  .LBB12_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v55, v57, v41, s4
; VI-NEXT:    v_perm_b32 v1, v1, v48, s4
; VI-NEXT:    v_lshlrev_b32_e32 v48, 16, v55
; VI-NEXT:    v_or_b32_e32 v1, v1, v48
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v34, v56, s4
; VI-NEXT:    v_perm_b32 v2, v2, v53, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v47, v40, s4
; VI-NEXT:    v_perm_b32 v2, v3, v38, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v46, v45, s4
; VI-NEXT:    v_perm_b32 v1, v4, v52, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v33, v39, s4
; VI-NEXT:    v_perm_b32 v2, v5, v37, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v44, v43, s4
; VI-NEXT:    v_perm_b32 v1, v6, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v50, v54, s4
; VI-NEXT:    v_perm_b32 v2, v7, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v62, v61, s4
; VI-NEXT:    v_perm_b32 v1, v8, v36, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v9, v49, s4
; VI-NEXT:    v_perm_b32 v1, v60, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v59, v58, s4
; VI-NEXT:    v_perm_b32 v1, v10, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v11, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v12, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v13, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v14, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v15, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v16, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v17, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v18, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v19, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v20, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v21, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v22, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v23, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v24, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v25, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v26, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v27, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v28, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v29, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v30, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v31, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v32, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v42, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32i32_to_v128i8:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(39)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB12_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(40)
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v7
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v1
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; GFX9-NEXT:  .LBB12_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB12_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    v_add_u32_e32 v32, 3, v32
; GFX9-NEXT:    s_waitcnt vmcnt(38)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6
; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5
; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4
; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3
; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v7
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v1
; GFX9-NEXT:  .LBB12_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v11, v11, v36, s4
; GFX9-NEXT:    v_perm_b32 v12, v12, v35, s4
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v3, v3, v34, s4
; GFX9-NEXT:    v_perm_b32 v34, v38, v39, s4
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v4, v4, v51, s4
; GFX9-NEXT:    v_perm_b32 v6, v6, v33, s4
; GFX9-NEXT:    v_perm_b32 v33, v50, v54, s4
; GFX9-NEXT:    v_perm_b32 v55, v58, v41, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v59, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, v2, v52, s4
; GFX9-NEXT:    v_perm_b32 v37, v37, v53, s4
; GFX9-NEXT:    v_perm_b32 v52, v57, v40, s4
; GFX9-NEXT:    v_perm_b32 v5, v5, v56, s4
; GFX9-NEXT:    v_perm_b32 v7, v7, v47, s4
; GFX9-NEXT:    v_perm_b32 v8, v8, v46, s4
; GFX9-NEXT:    v_perm_b32 v9, v9, v45, s4
; GFX9-NEXT:    v_perm_b32 v10, v10, v44, s4
; GFX9-NEXT:    v_perm_b32 v13, v13, v60, s4
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v35, v36, v35, s4
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v38, v63, v38, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v15, v15, v39, s4
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v14, v14, v36, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v16, v16, v39, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v36, v36, v50, s4
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v39, v39, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v17, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v50, v50, v53, s4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX9-NEXT:    v_perm_b32 v1, v49, v43, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshl_or_b32 v1, v52, 16, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    v_perm_b32 v1, v48, v62, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshl_or_b32 v1, v34, 16, v5
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v61, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v6
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshl_or_b32 v1, v33, 16, v7
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v8
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshl_or_b32 v1, v37, 16, v9
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v10
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshl_or_b32 v1, v38, 16, v11
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v12
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshl_or_b32 v1, v35, 16, v13
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v14
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshl_or_b32 v1, v36, 16, v15
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v16
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshl_or_b32 v1, v39, 16, v17
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v18
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v19, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v50, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v20, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v21, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v22, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v23, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v24, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v25, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v26, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v27, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v28, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v30, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v31, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v32, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v42, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v32i32_to_v128i8:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x2
; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB12_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB12_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB12_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 3, v32
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB12_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v1.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, v3.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v66, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v3, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, v4.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v65, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v6.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v39, 16, v2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v5.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, v8.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v7.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v66, v66, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v147, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v54, v64, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v39, 16, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v9.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v64, v65, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v66, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v55, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v10.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v11.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v64, 16, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v54, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v52, v55, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v12.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v13.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v14.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v15.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v12, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v13, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v14, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v15, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v16.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v17.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v18.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v19.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v16, v117, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v17, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v18, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v19, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v20.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v21.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v22.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v23.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v20, v103, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v21, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v99, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v14, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v24.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v15, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v25.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v26.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v27.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v27, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v28.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v29.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v30.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v31.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v28, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v29, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v30, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v31, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v32, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v29
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v32i32_to_v128i8:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x2
; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:12
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB12_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:  .LBB12_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB12_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 3, v32
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:  .LBB12_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v74, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v73, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v72, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v61, v65, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v59, v58, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v43, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v55, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v57, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v56, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v64, v46, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v41, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v182, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v53, v180, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v177, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v55, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v64, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v65, 16, v8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v54, 16, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v13, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v164, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v14, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v162, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v15, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v151, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v16, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v149, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v17, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v146, v49, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v18, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v144, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v19, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v133, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v20, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v131, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v21, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v128, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v118, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v115, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v24, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v113, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v25, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v102, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v26, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v97, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v28, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v87, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v29, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v84, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v30, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v31, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v71, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v69, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:88
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define inreg <128 x i8> @bitcast_v32i32_to_v128i8_scalar(<32 x i32> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32i32_to_v128i8_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_writelane_b32 v20, s30, 0
; SI-NEXT:    v_writelane_b32 v20, s31, 1
; SI-NEXT:    v_writelane_b32 v20, s34, 2
; SI-NEXT:    v_writelane_b32 v20, s35, 3
; SI-NEXT:    v_writelane_b32 v20, s36, 4
; SI-NEXT:    v_writelane_b32 v20, s37, 5
; SI-NEXT:    v_writelane_b32 v20, s38, 6
; SI-NEXT:    v_writelane_b32 v20, s39, 7
; SI-NEXT:    v_writelane_b32 v20, s48, 8
; SI-NEXT:    v_writelane_b32 v20, s49, 9
; SI-NEXT:    v_writelane_b32 v20, s50, 10
; SI-NEXT:    v_writelane_b32 v20, s51, 11
; SI-NEXT:    v_writelane_b32 v20, s52, 12
; SI-NEXT:    v_writelane_b32 v20, s53, 13
; SI-NEXT:    v_writelane_b32 v20, s54, 14
; SI-NEXT:    v_writelane_b32 v20, s55, 15
; SI-NEXT:    v_writelane_b32 v20, s64, 16
; SI-NEXT:    v_writelane_b32 v20, s65, 17
; SI-NEXT:    v_writelane_b32 v20, s66, 18
; SI-NEXT:    v_writelane_b32 v20, s67, 19
; SI-NEXT:    v_writelane_b32 v20, s68, 20
; SI-NEXT:    v_writelane_b32 v20, s69, 21
; SI-NEXT:    v_writelane_b32 v20, s70, 22
; SI-NEXT:    v_writelane_b32 v20, s71, 23
; SI-NEXT:    v_writelane_b32 v20, s80, 24
; SI-NEXT:    v_writelane_b32 v20, s81, 25
; SI-NEXT:    v_writelane_b32 v20, s82, 26
; SI-NEXT:    v_writelane_b32 v20, s83, 27
; SI-NEXT:    v_writelane_b32 v20, s84, 28
; SI-NEXT:    v_writelane_b32 v20, s85, 29
; SI-NEXT:    v_writelane_b32 v20, s86, 30
; SI-NEXT:    v_writelane_b32 v20, s87, 31
; SI-NEXT:    v_writelane_b32 v20, s96, 32
; SI-NEXT:    v_writelane_b32 v20, s97, 33
; SI-NEXT:    v_readfirstlane_b32 s44, v19
; SI-NEXT:    v_writelane_b32 v20, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s5, v18
; SI-NEXT:    v_readfirstlane_b32 s4, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s6, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s8, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s10, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s12, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s14, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s40, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s42, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v1
; SI-NEXT:    v_writelane_b32 v20, s99, 35
; SI-NEXT:    ; implicit-def: $vgpr22 : SGPR spill to VGPR lane
; SI-NEXT:    ; implicit-def: $vgpr21 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB13_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s46, s5, 24
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v22, s46, 42
; SI-NEXT:    s_lshr_b32 s46, s5, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 43
; SI-NEXT:    s_lshr_b32 s46, s5, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 44
; SI-NEXT:    s_lshr_b32 s46, s7, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 45
; SI-NEXT:    s_lshr_b32 s46, s7, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 46
; SI-NEXT:    s_lshr_b32 s46, s7, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 47
; SI-NEXT:    s_lshr_b32 s46, s9, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 48
; SI-NEXT:    s_lshr_b32 s46, s9, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 49
; SI-NEXT:    s_lshr_b32 s46, s11, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 50
; SI-NEXT:    s_lshr_b32 s46, s11, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 51
; SI-NEXT:    s_lshr_b32 s46, s11, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 52
; SI-NEXT:    s_lshr_b32 s46, s13, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 53
; SI-NEXT:    s_lshr_b32 s46, s13, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 54
; SI-NEXT:    s_lshr_b32 s46, s13, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 55
; SI-NEXT:    s_lshr_b32 s46, s15, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 56
; SI-NEXT:    s_lshr_b32 s46, s15, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 57
; SI-NEXT:    s_lshr_b32 s46, s15, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 58
; SI-NEXT:    s_lshr_b32 s46, s41, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 59
; SI-NEXT:    s_lshr_b32 s46, s41, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 60
; SI-NEXT:    s_lshr_b32 s46, s41, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 61
; SI-NEXT:    s_lshr_b32 s46, s43, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 62
; SI-NEXT:    s_lshr_b32 s46, s43, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 63
; SI-NEXT:    s_lshr_b32 s46, s43, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 0
; SI-NEXT:    s_lshr_b32 s46, s45, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 1
; SI-NEXT:    s_lshr_b32 s46, s45, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 2
; SI-NEXT:    s_lshr_b32 s46, s45, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 3
; SI-NEXT:    s_lshr_b32 s46, s29, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 4
; SI-NEXT:    s_lshr_b32 s46, s29, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 5
; SI-NEXT:    s_lshr_b32 s46, s29, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 6
; SI-NEXT:    s_lshr_b32 s46, s27, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 7
; SI-NEXT:    s_lshr_b32 s46, s27, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 8
; SI-NEXT:    s_lshr_b32 s46, s27, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 9
; SI-NEXT:    s_lshr_b32 s46, s25, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 10
; SI-NEXT:    s_lshr_b32 s46, s25, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 11
; SI-NEXT:    s_lshr_b32 s46, s25, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 12
; SI-NEXT:    s_lshr_b32 s46, s23, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 13
; SI-NEXT:    s_lshr_b32 s46, s23, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 14
; SI-NEXT:    s_lshr_b32 s46, s23, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 15
; SI-NEXT:    s_lshr_b32 s46, s21, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 16
; SI-NEXT:    s_lshr_b32 s46, s21, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 40
; SI-NEXT:    v_writelane_b32 v22, s47, 41
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 38
; SI-NEXT:    v_writelane_b32 v22, s47, 39
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 36
; SI-NEXT:    v_writelane_b32 v22, s47, 37
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 34
; SI-NEXT:    v_writelane_b32 v22, s47, 35
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 32
; SI-NEXT:    v_writelane_b32 v22, s47, 33
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 30
; SI-NEXT:    v_writelane_b32 v22, s47, 31
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 28
; SI-NEXT:    v_writelane_b32 v22, s47, 29
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 26
; SI-NEXT:    v_writelane_b32 v22, s47, 27
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 24
; SI-NEXT:    v_writelane_b32 v22, s47, 25
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 22
; SI-NEXT:    v_writelane_b32 v22, s47, 23
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 20
; SI-NEXT:    v_writelane_b32 v22, s47, 21
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 18
; SI-NEXT:    v_writelane_b32 v22, s47, 19
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 16
; SI-NEXT:    v_writelane_b32 v22, s47, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 14
; SI-NEXT:    v_writelane_b32 v22, s47, 15
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 12
; SI-NEXT:    v_writelane_b32 v22, s47, 13
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 10
; SI-NEXT:    v_writelane_b32 v22, s47, 11
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 8
; SI-NEXT:    v_writelane_b32 v22, s47, 9
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 6
; SI-NEXT:    v_writelane_b32 v22, s47, 7
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 4
; SI-NEXT:    v_writelane_b32 v22, s47, 5
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 2
; SI-NEXT:    v_writelane_b32 v22, s47, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 0
; SI-NEXT:    s_lshr_b32 s50, s9, 24
; SI-NEXT:    s_lshr_b32 s51, s21, 8
; SI-NEXT:    s_lshr_b32 s48, s19, 24
; SI-NEXT:    s_lshr_b32 s52, s19, 16
; SI-NEXT:    s_lshr_b32 s53, s19, 8
; SI-NEXT:    s_lshr_b32 s54, s17, 24
; SI-NEXT:    s_lshr_b32 s55, s17, 16
; SI-NEXT:    s_lshr_b32 s49, s17, 8
; SI-NEXT:    v_writelane_b32 v22, s47, 1
; SI-NEXT:    s_lshr_b64 s[64:65], s[42:43], 24
; SI-NEXT:    s_lshr_b64 s[66:67], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[68:69], s[42:43], 8
; SI-NEXT:    s_lshr_b64 s[70:71], s[44:45], 24
; SI-NEXT:    s_lshr_b64 s[80:81], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[84:85], s[28:29], 24
; SI-NEXT:    s_lshr_b64 s[86:87], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[96:97], s[28:29], 8
; SI-NEXT:    s_lshr_b64 s[98:99], s[26:27], 24
; SI-NEXT:    s_lshr_b64 s[46:47], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[26:27], 8
; SI-NEXT:    s_lshr_b64 s[58:59], s[24:25], 24
; SI-NEXT:    s_lshr_b64 s[60:61], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[24:25], 8
; SI-NEXT:    s_lshr_b64 s[72:73], s[22:23], 24
; SI-NEXT:    s_lshr_b64 s[74:75], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[22:23], 8
; SI-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; SI-NEXT:    s_lshr_b64 s[88:89], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[20:21], 8
; SI-NEXT:    s_lshr_b64 s[92:93], s[18:19], 24
; SI-NEXT:    s_lshr_b64 s[94:95], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[18:19], 8
; SI-NEXT:    s_lshr_b64 s[34:35], s[16:17], 24
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_lshr_b64 s[38:39], s[16:17], 8
; SI-NEXT:    s_cbranch_execnz .LBB13_3
; SI-NEXT:  .LBB13_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s5, s5, 3
; SI-NEXT:    s_add_i32 s4, s4, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 40
; SI-NEXT:    v_writelane_b32 v22, s47, 41
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 38
; SI-NEXT:    v_writelane_b32 v22, s47, 39
; SI-NEXT:    s_lshr_b32 s46, s5, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 42
; SI-NEXT:    s_lshr_b32 s46, s5, 16
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 43
; SI-NEXT:    s_lshr_b32 s46, s5, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 44
; SI-NEXT:    s_lshr_b32 s46, s7, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 45
; SI-NEXT:    s_lshr_b32 s46, s7, 16
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 46
; SI-NEXT:    s_lshr_b32 s46, s7, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 47
; SI-NEXT:    s_lshr_b32 s46, s9, 16
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 48
; SI-NEXT:    s_lshr_b32 s46, s9, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 49
; SI-NEXT:    s_lshr_b32 s46, s11, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 50
; SI-NEXT:    s_lshr_b32 s46, s11, 16
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 51
; SI-NEXT:    s_lshr_b32 s46, s11, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 52
; SI-NEXT:    s_lshr_b32 s46, s13, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 53
; SI-NEXT:    s_lshr_b32 s46, s13, 16
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 54
; SI-NEXT:    s_lshr_b32 s46, s13, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 55
; SI-NEXT:    s_lshr_b32 s46, s15, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 56
; SI-NEXT:    s_lshr_b32 s46, s15, 16
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 57
; SI-NEXT:    s_lshr_b32 s46, s15, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 58
; SI-NEXT:    s_lshr_b32 s46, s41, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 59
; SI-NEXT:    s_lshr_b32 s46, s41, 16
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 60
; SI-NEXT:    s_lshr_b32 s46, s41, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 61
; SI-NEXT:    s_lshr_b32 s46, s43, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 62
; SI-NEXT:    s_lshr_b32 s46, s43, 16
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 63
; SI-NEXT:    s_lshr_b32 s46, s43, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 0
; SI-NEXT:    s_lshr_b32 s46, s45, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 1
; SI-NEXT:    s_lshr_b32 s46, s45, 16
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    v_writelane_b32 v21, s46, 2
; SI-NEXT:    s_lshr_b32 s46, s45, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 3
; SI-NEXT:    s_lshr_b32 s46, s29, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 4
; SI-NEXT:    s_lshr_b32 s46, s29, 16
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    v_writelane_b32 v21, s46, 5
; SI-NEXT:    s_lshr_b32 s46, s29, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 6
; SI-NEXT:    s_lshr_b32 s46, s27, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 7
; SI-NEXT:    s_lshr_b32 s46, s27, 16
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    v_writelane_b32 v21, s46, 8
; SI-NEXT:    s_lshr_b32 s46, s27, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 9
; SI-NEXT:    s_lshr_b32 s46, s25, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 10
; SI-NEXT:    s_lshr_b32 s46, s25, 16
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    v_writelane_b32 v21, s46, 11
; SI-NEXT:    s_lshr_b32 s46, s25, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 12
; SI-NEXT:    s_lshr_b32 s46, s23, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 13
; SI-NEXT:    s_lshr_b32 s46, s23, 16
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    v_writelane_b32 v21, s46, 14
; SI-NEXT:    s_lshr_b32 s46, s23, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 15
; SI-NEXT:    s_lshr_b32 s46, s21, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 16
; SI-NEXT:    s_lshr_b32 s46, s21, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 8
; SI-NEXT:    s_add_i32 s6, s6, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 36
; SI-NEXT:    v_writelane_b32 v22, s47, 37
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 34
; SI-NEXT:    v_writelane_b32 v22, s47, 35
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 32
; SI-NEXT:    v_writelane_b32 v22, s47, 33
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 8
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 30
; SI-NEXT:    v_writelane_b32 v22, s47, 31
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 28
; SI-NEXT:    v_writelane_b32 v22, s47, 29
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 26
; SI-NEXT:    v_writelane_b32 v22, s47, 27
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 8
; SI-NEXT:    s_add_i32 s10, s10, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 24
; SI-NEXT:    v_writelane_b32 v22, s47, 25
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 22
; SI-NEXT:    v_writelane_b32 v22, s47, 23
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 20
; SI-NEXT:    v_writelane_b32 v22, s47, 21
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 8
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 18
; SI-NEXT:    v_writelane_b32 v22, s47, 19
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 16
; SI-NEXT:    v_writelane_b32 v22, s47, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 14
; SI-NEXT:    v_writelane_b32 v22, s47, 15
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 8
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 12
; SI-NEXT:    v_writelane_b32 v22, s47, 13
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 10
; SI-NEXT:    v_writelane_b32 v22, s47, 11
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 8
; SI-NEXT:    v_writelane_b32 v22, s47, 9
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 8
; SI-NEXT:    s_add_i32 s40, s40, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 6
; SI-NEXT:    v_writelane_b32 v22, s47, 7
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 4
; SI-NEXT:    v_writelane_b32 v22, s47, 5
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 2
; SI-NEXT:    v_writelane_b32 v22, s47, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 8
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s44, s44, 3
; SI-NEXT:    s_add_i32 s42, s42, 3
; SI-NEXT:    v_writelane_b32 v22, s46, 0
; SI-NEXT:    s_lshr_b32 s50, s9, 24
; SI-NEXT:    s_lshr_b32 s51, s21, 8
; SI-NEXT:    s_lshr_b32 s48, s19, 24
; SI-NEXT:    s_lshr_b32 s52, s19, 16
; SI-NEXT:    s_lshr_b32 s53, s19, 8
; SI-NEXT:    s_lshr_b32 s54, s17, 24
; SI-NEXT:    s_lshr_b32 s55, s17, 16
; SI-NEXT:    s_lshr_b32 s49, s17, 8
; SI-NEXT:    v_writelane_b32 v22, s47, 1
; SI-NEXT:    s_lshr_b64 s[64:65], s[42:43], 24
; SI-NEXT:    s_lshr_b64 s[66:67], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[68:69], s[42:43], 8
; SI-NEXT:    s_lshr_b64 s[70:71], s[44:45], 24
; SI-NEXT:    s_lshr_b64 s[80:81], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[84:85], s[28:29], 24
; SI-NEXT:    s_lshr_b64 s[86:87], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[96:97], s[28:29], 8
; SI-NEXT:    s_lshr_b64 s[98:99], s[26:27], 24
; SI-NEXT:    s_lshr_b64 s[46:47], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[26:27], 8
; SI-NEXT:    s_lshr_b64 s[58:59], s[24:25], 24
; SI-NEXT:    s_lshr_b64 s[60:61], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[24:25], 8
; SI-NEXT:    s_lshr_b64 s[72:73], s[22:23], 24
; SI-NEXT:    s_lshr_b64 s[74:75], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[22:23], 8
; SI-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; SI-NEXT:    s_lshr_b64 s[88:89], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[20:21], 8
; SI-NEXT:    s_lshr_b64 s[92:93], s[18:19], 24
; SI-NEXT:    s_lshr_b64 s[94:95], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[18:19], 8
; SI-NEXT:    s_lshr_b64 s[34:35], s[16:17], 24
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_lshr_b64 s[38:39], s[16:17], 8
; SI-NEXT:  .LBB13_3: ; %end
; SI-NEXT:    s_lshl_b32 s47, s38, 8
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    s_and_b32 s47, s36, 0xff
; SI-NEXT:    s_lshl_b32 s57, s34, 24
; SI-NEXT:    s_lshl_b32 s47, s47, 16
; SI-NEXT:    s_or_b32 s47, s57, s47
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    v_mov_b32_e32 v1, s16
; SI-NEXT:    s_and_b32 s16, s17, 0xff
; SI-NEXT:    s_lshl_b32 s17, s49, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s55, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s47, s54, 24
; SI-NEXT:    s_or_b32 s17, s47, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_lshl_b32 s16, s30, 8
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s94, 0xff
; SI-NEXT:    s_lshl_b32 s18, s92, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v3, s16
; SI-NEXT:    s_and_b32 s16, s19, 0xff
; SI-NEXT:    s_lshl_b32 s17, s53, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s52, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s48, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v4, s16
; SI-NEXT:    s_lshl_b32 s16, s90, 8
; SI-NEXT:    s_and_b32 s17, s20, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s88, 0xff
; SI-NEXT:    s_lshl_b32 s18, s78, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v5, s16
; SI-NEXT:    s_and_b32 s16, s21, 0xff
; SI-NEXT:    s_lshl_b32 s17, s51, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 17
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 16
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v6, s16
; SI-NEXT:    s_lshl_b32 s16, s76, 8
; SI-NEXT:    s_and_b32 s17, s22, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s74, 0xff
; SI-NEXT:    s_lshl_b32 s18, s72, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 15
; SI-NEXT:    v_mov_b32_e32 v7, s16
; SI-NEXT:    s_and_b32 s16, s23, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 14
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 13
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v8, s16
; SI-NEXT:    s_lshl_b32 s16, s62, 8
; SI-NEXT:    s_and_b32 s17, s24, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s60, 0xff
; SI-NEXT:    s_lshl_b32 s18, s58, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 12
; SI-NEXT:    v_mov_b32_e32 v9, s16
; SI-NEXT:    s_and_b32 s16, s25, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 11
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 10
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v10, s16
; SI-NEXT:    s_lshl_b32 s16, s56, 8
; SI-NEXT:    s_and_b32 s17, s26, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s46, 0xff
; SI-NEXT:    s_lshl_b32 s18, s98, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 9
; SI-NEXT:    v_mov_b32_e32 v11, s16
; SI-NEXT:    s_and_b32 s16, s27, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 8
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 7
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v12, s16
; SI-NEXT:    s_lshl_b32 s16, s96, 8
; SI-NEXT:    s_and_b32 s17, s28, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s86, 0xff
; SI-NEXT:    s_lshl_b32 s18, s84, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 6
; SI-NEXT:    v_mov_b32_e32 v13, s16
; SI-NEXT:    s_and_b32 s16, s29, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 5
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 4
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v14, s16
; SI-NEXT:    s_lshl_b32 s16, s82, 8
; SI-NEXT:    s_and_b32 s17, s44, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s80, 0xff
; SI-NEXT:    s_lshl_b32 s18, s70, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 3
; SI-NEXT:    v_mov_b32_e32 v15, s16
; SI-NEXT:    s_and_b32 s16, s45, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 2
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 1
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 4, v0
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 12, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v4, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 16, v0
; SI-NEXT:    v_mov_b32_e32 v16, s16
; SI-NEXT:    s_lshl_b32 s16, s68, 8
; SI-NEXT:    s_and_b32 s17, s42, 0xff
; SI-NEXT:    buffer_store_dword v5, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 20, v0
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s66, 0xff
; SI-NEXT:    buffer_store_dword v6, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 24, v0
; SI-NEXT:    s_lshl_b32 s18, s64, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    buffer_store_dword v7, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 28, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    buffer_store_dword v8, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 36, v0
; SI-NEXT:    v_readlane_b32 s17, v21, 0
; SI-NEXT:    buffer_store_dword v10, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 40, v0
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s43, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    buffer_store_dword v11, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 44, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 63
; SI-NEXT:    buffer_store_dword v12, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 48, v0
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 62
; SI-NEXT:    buffer_store_dword v13, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 52, v0
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    buffer_store_dword v14, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 56, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_readlane_b32 s18, v22, 0
; SI-NEXT:    buffer_store_dword v15, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 60, v0
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s19, v22, 1
; SI-NEXT:    buffer_store_dword v16, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 64, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_lshl_b32 s17, s18, 8
; SI-NEXT:    v_readlane_b32 s18, v22, 2
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s40, 0xff
; SI-NEXT:    v_readlane_b32 s19, v22, 3
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 4
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x44, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 61
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s41, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 60
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 59
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x48, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    v_readlane_b32 s16, v22, 6
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s17, v22, 7
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    v_readlane_b32 s19, v22, 5
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    v_readlane_b32 s16, v22, 8
; SI-NEXT:    v_readlane_b32 s17, v22, 9
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 10
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s17, s18, 24
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x4c, v0
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s14
; SI-NEXT:    s_and_b32 s14, s15, 0xff
; SI-NEXT:    v_readlane_b32 s15, v22, 58
; SI-NEXT:    s_lshl_b32 s15, s15, 8
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    v_readlane_b32 s15, v22, 57
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    v_readlane_b32 s16, v22, 56
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_lshl_b32 s16, s16, 24
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x50, v0
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s14
; SI-NEXT:    v_readlane_b32 s14, v22, 12
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s15, v22, 13
; SI-NEXT:    s_lshl_b32 s14, s14, 8
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    v_readlane_b32 s14, v22, 14
; SI-NEXT:    v_readlane_b32 s15, v22, 15
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s16, v22, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_lshl_b32 s15, s16, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x54, v0
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s12
; SI-NEXT:    s_and_b32 s12, s13, 0xff
; SI-NEXT:    v_readlane_b32 s13, v22, 55
; SI-NEXT:    s_lshl_b32 s13, s13, 8
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    v_readlane_b32 s13, v22, 54
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    v_readlane_b32 s14, v22, 53
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x58, v0
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s12
; SI-NEXT:    v_readlane_b32 s12, v22, 18
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s13, v22, 19
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    v_readlane_b32 s12, v22, 20
; SI-NEXT:    v_readlane_b32 s13, v22, 21
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s14, v22, 22
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s13, s14, 24
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x5c, v0
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s10
; SI-NEXT:    s_and_b32 s10, s11, 0xff
; SI-NEXT:    v_readlane_b32 s11, v22, 52
; SI-NEXT:    s_lshl_b32 s11, s11, 8
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    v_readlane_b32 s11, v22, 51
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    v_readlane_b32 s12, v22, 50
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_lshl_b32 s12, s12, 24
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x60, v0
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s10
; SI-NEXT:    v_readlane_b32 s10, v22, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s11, v22, 25
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    v_readlane_b32 s10, v22, 26
; SI-NEXT:    v_readlane_b32 s11, v22, 27
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s12, v22, 28
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s11, s12, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x64, v0
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    s_and_b32 s8, s9, 0xff
; SI-NEXT:    v_readlane_b32 s9, v22, 49
; SI-NEXT:    s_lshl_b32 s9, s9, 8
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    v_readlane_b32 s9, v22, 48
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_lshl_b32 s10, s50, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x68, v0
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    v_readlane_b32 s8, v22, 30
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s9, v22, 31
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    v_readlane_b32 s8, v22, 32
; SI-NEXT:    v_readlane_b32 s9, v22, 33
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v22, 34
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s10, 24
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x6c, v0
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    s_and_b32 s6, s7, 0xff
; SI-NEXT:    v_readlane_b32 s7, v22, 47
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    v_readlane_b32 s7, v22, 46
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    v_readlane_b32 s8, v22, 45
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x70, v0
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_readlane_b32 s6, v22, 36
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    v_readlane_b32 s7, v22, 37
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    v_readlane_b32 s6, v22, 38
; SI-NEXT:    v_readlane_b32 s7, v22, 39
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s8, v22, 40
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s8, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x74, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    s_and_b32 s4, s5, 0xff
; SI-NEXT:    v_readlane_b32 s5, v22, 44
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_readlane_b32 s5, v22, 43
; SI-NEXT:    s_and_b32 s5, s5, 0xff
; SI-NEXT:    v_readlane_b32 s6, v22, 42
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x78, v0
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    v_mov_b32_e32 v1, s4
; SI-NEXT:    v_readlane_b32 s19, v22, 11
; SI-NEXT:    v_readlane_b32 s17, v22, 17
; SI-NEXT:    v_readlane_b32 s15, v22, 23
; SI-NEXT:    v_readlane_b32 s13, v22, 29
; SI-NEXT:    v_readlane_b32 s11, v22, 35
; SI-NEXT:    v_readlane_b32 s9, v22, 41
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    v_readlane_b32 s99, v20, 35
; SI-NEXT:    v_readlane_b32 s98, v20, 34
; SI-NEXT:    v_readlane_b32 s97, v20, 33
; SI-NEXT:    v_readlane_b32 s96, v20, 32
; SI-NEXT:    v_readlane_b32 s87, v20, 31
; SI-NEXT:    v_readlane_b32 s86, v20, 30
; SI-NEXT:    v_readlane_b32 s85, v20, 29
; SI-NEXT:    v_readlane_b32 s84, v20, 28
; SI-NEXT:    v_readlane_b32 s83, v20, 27
; SI-NEXT:    v_readlane_b32 s82, v20, 26
; SI-NEXT:    v_readlane_b32 s81, v20, 25
; SI-NEXT:    v_readlane_b32 s80, v20, 24
; SI-NEXT:    v_readlane_b32 s71, v20, 23
; SI-NEXT:    v_readlane_b32 s70, v20, 22
; SI-NEXT:    v_readlane_b32 s69, v20, 21
; SI-NEXT:    v_readlane_b32 s68, v20, 20
; SI-NEXT:    v_readlane_b32 s67, v20, 19
; SI-NEXT:    v_readlane_b32 s66, v20, 18
; SI-NEXT:    v_readlane_b32 s65, v20, 17
; SI-NEXT:    v_readlane_b32 s64, v20, 16
; SI-NEXT:    v_readlane_b32 s55, v20, 15
; SI-NEXT:    v_readlane_b32 s54, v20, 14
; SI-NEXT:    v_readlane_b32 s53, v20, 13
; SI-NEXT:    v_readlane_b32 s52, v20, 12
; SI-NEXT:    v_readlane_b32 s51, v20, 11
; SI-NEXT:    v_readlane_b32 s50, v20, 10
; SI-NEXT:    v_readlane_b32 s49, v20, 9
; SI-NEXT:    v_readlane_b32 s48, v20, 8
; SI-NEXT:    v_readlane_b32 s39, v20, 7
; SI-NEXT:    v_readlane_b32 s38, v20, 6
; SI-NEXT:    v_readlane_b32 s37, v20, 5
; SI-NEXT:    v_readlane_b32 s36, v20, 4
; SI-NEXT:    v_readlane_b32 s35, v20, 3
; SI-NEXT:    v_readlane_b32 s34, v20, 2
; SI-NEXT:    v_readlane_b32 s31, v20, 1
; SI-NEXT:    v_readlane_b32 s30, v20, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB13_4:
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v22, s50, 0
; SI-NEXT:    v_writelane_b32 v22, s51, 1
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 2
; SI-NEXT:    v_writelane_b32 v22, s51, 3
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 4
; SI-NEXT:    v_writelane_b32 v22, s51, 5
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 6
; SI-NEXT:    v_writelane_b32 v22, s51, 7
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 8
; SI-NEXT:    v_writelane_b32 v22, s51, 9
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 10
; SI-NEXT:    v_writelane_b32 v22, s51, 11
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 12
; SI-NEXT:    v_writelane_b32 v22, s51, 13
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 14
; SI-NEXT:    v_writelane_b32 v22, s51, 15
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 16
; SI-NEXT:    v_writelane_b32 v22, s51, 17
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 18
; SI-NEXT:    v_writelane_b32 v22, s51, 19
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 20
; SI-NEXT:    v_writelane_b32 v22, s51, 21
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 22
; SI-NEXT:    v_writelane_b32 v22, s51, 23
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 24
; SI-NEXT:    v_writelane_b32 v22, s51, 25
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 26
; SI-NEXT:    v_writelane_b32 v22, s51, 27
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 28
; SI-NEXT:    v_writelane_b32 v22, s51, 29
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 30
; SI-NEXT:    v_writelane_b32 v22, s51, 31
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 32
; SI-NEXT:    v_writelane_b32 v22, s51, 33
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 34
; SI-NEXT:    v_writelane_b32 v22, s51, 35
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 36
; SI-NEXT:    v_writelane_b32 v22, s51, 37
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 38
; SI-NEXT:    v_writelane_b32 v22, s51, 39
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s50, 40
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    v_writelane_b32 v22, s51, 41
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    s_branch .LBB13_2
;
; VI-LABEL: bitcast_v32i32_to_v128i8_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s34, 2
; VI-NEXT:    v_writelane_b32 v32, s35, 3
; VI-NEXT:    v_writelane_b32 v32, s36, 4
; VI-NEXT:    v_writelane_b32 v32, s37, 5
; VI-NEXT:    v_writelane_b32 v32, s38, 6
; VI-NEXT:    v_writelane_b32 v32, s39, 7
; VI-NEXT:    v_writelane_b32 v32, s48, 8
; VI-NEXT:    v_writelane_b32 v32, s49, 9
; VI-NEXT:    v_writelane_b32 v32, s50, 10
; VI-NEXT:    v_writelane_b32 v32, s51, 11
; VI-NEXT:    v_writelane_b32 v32, s52, 12
; VI-NEXT:    v_writelane_b32 v32, s53, 13
; VI-NEXT:    v_writelane_b32 v32, s54, 14
; VI-NEXT:    v_writelane_b32 v32, s55, 15
; VI-NEXT:    v_writelane_b32 v32, s64, 16
; VI-NEXT:    v_writelane_b32 v32, s65, 17
; VI-NEXT:    v_writelane_b32 v32, s66, 18
; VI-NEXT:    v_writelane_b32 v32, s67, 19
; VI-NEXT:    v_writelane_b32 v32, s68, 20
; VI-NEXT:    v_writelane_b32 v32, s69, 21
; VI-NEXT:    v_writelane_b32 v32, s70, 22
; VI-NEXT:    v_writelane_b32 v32, s71, 23
; VI-NEXT:    v_writelane_b32 v32, s80, 24
; VI-NEXT:    v_writelane_b32 v32, s81, 25
; VI-NEXT:    v_writelane_b32 v32, s82, 26
; VI-NEXT:    v_writelane_b32 v32, s83, 27
; VI-NEXT:    v_writelane_b32 v32, s84, 28
; VI-NEXT:    v_writelane_b32 v32, s85, 29
; VI-NEXT:    v_readfirstlane_b32 s44, v19
; VI-NEXT:    v_writelane_b32 v32, s86, 30
; VI-NEXT:    v_readfirstlane_b32 s5, v18
; VI-NEXT:    v_readfirstlane_b32 s4, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s6, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s8, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s10, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s12, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s14, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s40, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s42, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    s_cmp_lg_u32 s44, 0
; VI-NEXT:    v_readfirstlane_b32 s44, v1
; VI-NEXT:    v_writelane_b32 v32, s87, 31
; VI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; VI-NEXT:    s_cbranch_scc0 .LBB13_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s14, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s14, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s41, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s41, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s40, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s40, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s43, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s43, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s42, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s42, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 57
; VI-NEXT:    s_lshr_b64 s[56:57], s[4:5], 24
; VI-NEXT:    v_writelane_b32 v33, s56, 0
; VI-NEXT:    s_lshr_b32 s51, s29, 8
; VI-NEXT:    s_lshr_b32 s86, s28, 8
; VI-NEXT:    s_lshr_b32 s52, s27, 8
; VI-NEXT:    s_lshr_b32 s66, s26, 8
; VI-NEXT:    s_lshr_b32 s64, s25, 8
; VI-NEXT:    s_lshr_b32 s50, s24, 8
; VI-NEXT:    s_lshr_b32 s71, s23, 24
; VI-NEXT:    s_lshr_b32 s70, s23, 16
; VI-NEXT:    s_lshr_b32 s55, s23, 8
; VI-NEXT:    s_lshr_b32 s65, s22, 8
; VI-NEXT:    s_lshr_b32 s68, s21, 24
; VI-NEXT:    s_lshr_b32 s54, s21, 8
; VI-NEXT:    s_lshr_b32 s84, s20, 16
; VI-NEXT:    s_lshr_b32 s69, s20, 8
; VI-NEXT:    s_lshr_b32 s80, s19, 24
; VI-NEXT:    s_lshr_b32 s87, s19, 16
; VI-NEXT:    s_lshr_b32 s46, s19, 8
; VI-NEXT:    s_lshr_b32 s81, s18, 16
; VI-NEXT:    s_lshr_b32 s47, s18, 8
; VI-NEXT:    s_lshr_b32 s85, s17, 24
; VI-NEXT:    s_lshr_b32 s83, s17, 16
; VI-NEXT:    s_lshr_b32 s67, s17, 8
; VI-NEXT:    s_lshr_b32 s53, s16, 16
; VI-NEXT:    s_lshr_b32 s82, s16, 8
; VI-NEXT:    v_writelane_b32 v33, s57, 1
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:    s_cbranch_execnz .LBB13_3
; VI-NEXT:  .LBB13_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s5, s5, 3
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    s_add_i32 s4, s4, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s14, 16
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s14, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s41, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s41, 16
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s40, 16
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s40, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s43, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s43, 16
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s42, 16
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s42, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    v_writelane_b32 v33, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 57
; VI-NEXT:    s_lshr_b64 s[56:57], s[4:5], 24
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    v_writelane_b32 v33, s56, 0
; VI-NEXT:    s_lshr_b32 s51, s29, 8
; VI-NEXT:    s_lshr_b32 s86, s28, 8
; VI-NEXT:    s_lshr_b32 s52, s27, 8
; VI-NEXT:    s_lshr_b32 s66, s26, 8
; VI-NEXT:    s_lshr_b32 s64, s25, 8
; VI-NEXT:    s_lshr_b32 s50, s24, 8
; VI-NEXT:    s_lshr_b32 s71, s23, 24
; VI-NEXT:    s_lshr_b32 s70, s23, 16
; VI-NEXT:    s_lshr_b32 s55, s23, 8
; VI-NEXT:    s_lshr_b32 s65, s22, 8
; VI-NEXT:    s_lshr_b32 s68, s21, 24
; VI-NEXT:    s_lshr_b32 s54, s21, 8
; VI-NEXT:    s_lshr_b32 s84, s20, 16
; VI-NEXT:    s_lshr_b32 s69, s20, 8
; VI-NEXT:    s_lshr_b32 s80, s19, 24
; VI-NEXT:    s_lshr_b32 s87, s19, 16
; VI-NEXT:    s_lshr_b32 s46, s19, 8
; VI-NEXT:    s_lshr_b32 s81, s18, 16
; VI-NEXT:    s_lshr_b32 s47, s18, 8
; VI-NEXT:    s_lshr_b32 s85, s17, 24
; VI-NEXT:    s_lshr_b32 s83, s17, 16
; VI-NEXT:    s_lshr_b32 s67, s17, 8
; VI-NEXT:    s_lshr_b32 s53, s16, 16
; VI-NEXT:    s_lshr_b32 s82, s16, 8
; VI-NEXT:    v_writelane_b32 v33, s57, 1
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:  .LBB13_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v8, s48
; VI-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; VI-NEXT:    v_mov_b32_e32 v2, s82
; VI-NEXT:    v_perm_b32 v8, s53, v8, v1
; VI-NEXT:    v_perm_b32 v3, s16, v2, v1
; VI-NEXT:    v_mov_b32_e32 v26, s85
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v4, s67
; VI-NEXT:    v_perm_b32 v26, s83, v26, v1
; VI-NEXT:    v_or_b32_e32 v3, v3, v8
; VI-NEXT:    v_perm_b32 v2, s17, v4, v1
; VI-NEXT:    v_mov_b32_e32 v12, s38
; VI-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v26
; VI-NEXT:    v_mov_b32_e32 v4, s47
; VI-NEXT:    v_perm_b32 v12, s81, v12, v1
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 4, v0
; VI-NEXT:    v_perm_b32 v5, s18, v4, v1
; VI-NEXT:    v_mov_b32_e32 v27, s80
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v12
; VI-NEXT:    v_mov_b32_e32 v6, s46
; VI-NEXT:    v_perm_b32 v27, s87, v27, v1
; VI-NEXT:    v_or_b32_e32 v2, v5, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 8, v0
; VI-NEXT:    v_perm_b32 v4, s19, v6, v1
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_readlane_b32 s16, v33, 46
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
; VI-NEXT:    v_mov_b32_e32 v6, s69
; VI-NEXT:    v_mov_b32_e32 v24, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 44
; VI-NEXT:    v_perm_b32 v16, s84, v16, v1
; VI-NEXT:    v_or_b32_e32 v2, v4, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 12, v0
; VI-NEXT:    v_perm_b32 v7, s20, v6, v1
; VI-NEXT:    v_mov_b32_e32 v25, s16
; VI-NEXT:    v_mov_b32_e32 v28, s68
; VI-NEXT:    v_readlane_b32 s16, v33, 57
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v16
; VI-NEXT:    v_mov_b32_e32 v9, s54
; VI-NEXT:    v_perm_b32 v28, s16, v28, v1
; VI-NEXT:    v_or_b32_e32 v2, v7, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 16, v0
; VI-NEXT:    v_perm_b32 v6, s21, v9, v1
; VI-NEXT:    v_mov_b32_e32 v19, s34
; VI-NEXT:    v_readlane_b32 s16, v33, 56
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
; VI-NEXT:    v_mov_b32_e32 v9, s65
; VI-NEXT:    v_perm_b32 v19, s16, v19, v1
; VI-NEXT:    v_or_b32_e32 v2, v6, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 20, v0
; VI-NEXT:    v_perm_b32 v10, s22, v9, v1
; VI-NEXT:    v_mov_b32_e32 v29, s71
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v19
; VI-NEXT:    v_mov_b32_e32 v11, s55
; VI-NEXT:    v_perm_b32 v29, s70, v29, v1
; VI-NEXT:    v_or_b32_e32 v2, v10, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 24, v0
; VI-NEXT:    v_perm_b32 v9, s23, v11, v1
; VI-NEXT:    v_mov_b32_e32 v20, s30
; VI-NEXT:    v_readlane_b32 s16, v33, 55
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v29
; VI-NEXT:    v_mov_b32_e32 v11, s50
; VI-NEXT:    v_perm_b32 v20, s16, v20, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 53
; VI-NEXT:    v_or_b32_e32 v2, v9, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 28, v0
; VI-NEXT:    v_perm_b32 v13, s24, v11, v1
; VI-NEXT:    v_mov_b32_e32 v30, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 54
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v20
; VI-NEXT:    v_mov_b32_e32 v14, s64
; VI-NEXT:    v_perm_b32 v30, s16, v30, v1
; VI-NEXT:    v_or_b32_e32 v2, v13, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 32, v0
; VI-NEXT:    v_perm_b32 v11, s25, v14, v1
; VI-NEXT:    v_mov_b32_e32 v21, s90
; VI-NEXT:    v_readlane_b32 s16, v33, 52
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
; VI-NEXT:    v_mov_b32_e32 v14, s66
; VI-NEXT:    v_perm_b32 v21, s16, v21, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 50
; VI-NEXT:    v_or_b32_e32 v2, v11, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 36, v0
; VI-NEXT:    v_perm_b32 v15, s26, v14, v1
; VI-NEXT:    v_mov_b32_e32 v31, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 51
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v21
; VI-NEXT:    v_mov_b32_e32 v17, s52
; VI-NEXT:    v_perm_b32 v31, s16, v31, v1
; VI-NEXT:    v_or_b32_e32 v2, v15, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 40, v0
; VI-NEXT:    v_perm_b32 v14, s27, v17, v1
; VI-NEXT:    v_mov_b32_e32 v22, s88
; VI-NEXT:    v_readlane_b32 s16, v33, 49
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v31
; VI-NEXT:    v_mov_b32_e32 v17, s86
; VI-NEXT:    v_perm_b32 v22, s16, v22, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 47
; VI-NEXT:    v_or_b32_e32 v2, v14, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 44, v0
; VI-NEXT:    v_perm_b32 v18, s28, v17, v1
; VI-NEXT:    v_mov_b32_e32 v8, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 48
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v22
; VI-NEXT:    v_mov_b32_e32 v23, s51
; VI-NEXT:    v_perm_b32 v3, s16, v8, v1
; VI-NEXT:    v_or_b32_e32 v2, v18, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 48, v0
; VI-NEXT:    v_perm_b32 v17, s29, v23, v1
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; VI-NEXT:    v_mov_b32_e32 v23, s78
; VI-NEXT:    v_or_b32_e32 v2, v17, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 52, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 45
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, s16, v23, v1
; VI-NEXT:    v_perm_b32 v24, s44, v24, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v2, v24, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 56, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 42
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 43
; VI-NEXT:    v_perm_b32 v2, s16, v2, v1
; VI-NEXT:    v_perm_b32 v25, s45, v25, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v2, v25, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 60, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 41
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_mov_b32_e32 v3, s76
; VI-NEXT:    v_readlane_b32 s16, v33, 40
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s42, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 64, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 39
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 37
; VI-NEXT:    v_mov_b32_e32 v3, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 38
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s43, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x44, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 36
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_mov_b32_e32 v3, s74
; VI-NEXT:    v_readlane_b32 s16, v33, 35
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s40, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x48, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 34
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 32
; VI-NEXT:    v_mov_b32_e32 v3, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 33
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s41, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x4c, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 31
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_perm_b32 v2, s14, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s72
; VI-NEXT:    v_readlane_b32 s14, v33, 30
; VI-NEXT:    v_perm_b32 v3, s14, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x50, v0
; VI-NEXT:    v_readlane_b32 s14, v33, 29
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s14
; VI-NEXT:    v_readlane_b32 s14, v33, 27
; VI-NEXT:    v_mov_b32_e32 v3, s14
; VI-NEXT:    v_readlane_b32 s14, v33, 28
; VI-NEXT:    v_perm_b32 v3, s14, v3, v1
; VI-NEXT:    v_perm_b32 v2, s15, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x54, v0
; VI-NEXT:    v_readlane_b32 s14, v33, 26
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s14
; VI-NEXT:    v_perm_b32 v2, s12, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s62
; VI-NEXT:    v_readlane_b32 s12, v33, 25
; VI-NEXT:    v_perm_b32 v3, s12, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x58, v0
; VI-NEXT:    v_readlane_b32 s12, v33, 24
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s12
; VI-NEXT:    v_readlane_b32 s12, v33, 22
; VI-NEXT:    v_mov_b32_e32 v3, s12
; VI-NEXT:    v_readlane_b32 s12, v33, 23
; VI-NEXT:    v_perm_b32 v3, s12, v3, v1
; VI-NEXT:    v_perm_b32 v2, s13, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x5c, v0
; VI-NEXT:    v_readlane_b32 s12, v33, 21
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s12
; VI-NEXT:    v_perm_b32 v2, s10, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s60
; VI-NEXT:    v_readlane_b32 s10, v33, 20
; VI-NEXT:    v_perm_b32 v3, s10, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x60, v0
; VI-NEXT:    v_readlane_b32 s10, v33, 19
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s10
; VI-NEXT:    v_readlane_b32 s10, v33, 17
; VI-NEXT:    v_mov_b32_e32 v3, s10
; VI-NEXT:    v_readlane_b32 s10, v33, 18
; VI-NEXT:    v_perm_b32 v3, s10, v3, v1
; VI-NEXT:    v_perm_b32 v2, s11, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x64, v0
; VI-NEXT:    v_readlane_b32 s10, v33, 16
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s10
; VI-NEXT:    v_perm_b32 v2, s8, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s58
; VI-NEXT:    v_readlane_b32 s8, v33, 15
; VI-NEXT:    v_perm_b32 v3, s8, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x68, v0
; VI-NEXT:    v_readlane_b32 s8, v33, 14
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s8
; VI-NEXT:    v_readlane_b32 s8, v33, 12
; VI-NEXT:    v_mov_b32_e32 v3, s8
; VI-NEXT:    v_readlane_b32 s8, v33, 13
; VI-NEXT:    v_perm_b32 v3, s8, v3, v1
; VI-NEXT:    v_perm_b32 v2, s9, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x6c, v0
; VI-NEXT:    v_readlane_b32 s8, v33, 11
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s8
; VI-NEXT:    v_perm_b32 v2, s6, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s56
; VI-NEXT:    v_readlane_b32 s6, v33, 10
; VI-NEXT:    v_perm_b32 v3, s6, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x70, v0
; VI-NEXT:    v_readlane_b32 s6, v33, 9
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 7
; VI-NEXT:    v_mov_b32_e32 v3, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 8
; VI-NEXT:    v_perm_b32 v3, s6, v3, v1
; VI-NEXT:    v_perm_b32 v2, s7, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x74, v0
; VI-NEXT:    v_readlane_b32 s6, v33, 6
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 0
; VI-NEXT:    v_perm_b32 v2, s4, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s6
; VI-NEXT:    v_readlane_b32 s4, v33, 5
; VI-NEXT:    v_perm_b32 v3, s4, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x78, v0
; VI-NEXT:    v_readlane_b32 s4, v33, 4
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s4
; VI-NEXT:    v_readlane_b32 s4, v33, 2
; VI-NEXT:    v_mov_b32_e32 v3, s4
; VI-NEXT:    v_readlane_b32 s4, v33, 3
; VI-NEXT:    v_perm_b32 v2, s5, v2, v1
; VI-NEXT:    v_perm_b32 v1, s4, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    v_readlane_b32 s7, v33, 1
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_readlane_b32 s87, v32, 31
; VI-NEXT:    v_readlane_b32 s86, v32, 30
; VI-NEXT:    v_readlane_b32 s85, v32, 29
; VI-NEXT:    v_readlane_b32 s84, v32, 28
; VI-NEXT:    v_readlane_b32 s83, v32, 27
; VI-NEXT:    v_readlane_b32 s82, v32, 26
; VI-NEXT:    v_readlane_b32 s81, v32, 25
; VI-NEXT:    v_readlane_b32 s80, v32, 24
; VI-NEXT:    v_readlane_b32 s71, v32, 23
; VI-NEXT:    v_readlane_b32 s70, v32, 22
; VI-NEXT:    v_readlane_b32 s69, v32, 21
; VI-NEXT:    v_readlane_b32 s68, v32, 20
; VI-NEXT:    v_readlane_b32 s67, v32, 19
; VI-NEXT:    v_readlane_b32 s66, v32, 18
; VI-NEXT:    v_readlane_b32 s65, v32, 17
; VI-NEXT:    v_readlane_b32 s64, v32, 16
; VI-NEXT:    v_readlane_b32 s55, v32, 15
; VI-NEXT:    v_readlane_b32 s54, v32, 14
; VI-NEXT:    v_readlane_b32 s53, v32, 13
; VI-NEXT:    v_readlane_b32 s52, v32, 12
; VI-NEXT:    v_readlane_b32 s51, v32, 11
; VI-NEXT:    v_readlane_b32 s50, v32, 10
; VI-NEXT:    v_readlane_b32 s49, v32, 9
; VI-NEXT:    v_readlane_b32 s48, v32, 8
; VI-NEXT:    v_readlane_b32 s39, v32, 7
; VI-NEXT:    v_readlane_b32 s38, v32, 6
; VI-NEXT:    v_readlane_b32 s37, v32, 5
; VI-NEXT:    v_readlane_b32 s36, v32, 4
; VI-NEXT:    v_readlane_b32 s35, v32, 3
; VI-NEXT:    v_readlane_b32 s34, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB13_4:
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr71
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    v_writelane_b32 v33, s70, 0
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr82
; VI-NEXT:    ; implicit-def: $sgpr53
; VI-NEXT:    ; implicit-def: $sgpr67
; VI-NEXT:    ; implicit-def: $sgpr83
; VI-NEXT:    ; implicit-def: $sgpr85
; VI-NEXT:    ; implicit-def: $sgpr47
; VI-NEXT:    ; implicit-def: $sgpr81
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr87
; VI-NEXT:    ; implicit-def: $sgpr80
; VI-NEXT:    ; implicit-def: $sgpr69
; VI-NEXT:    ; implicit-def: $sgpr84
; VI-NEXT:    ; implicit-def: $sgpr54
; VI-NEXT:    ; implicit-def: $sgpr68
; VI-NEXT:    ; implicit-def: $sgpr65
; VI-NEXT:    ; implicit-def: $sgpr55
; VI-NEXT:    ; implicit-def: $sgpr50
; VI-NEXT:    ; implicit-def: $sgpr64
; VI-NEXT:    ; implicit-def: $sgpr66
; VI-NEXT:    ; implicit-def: $sgpr52
; VI-NEXT:    ; implicit-def: $sgpr86
; VI-NEXT:    ; implicit-def: $sgpr51
; VI-NEXT:    ; implicit-def: $sgpr48
; VI-NEXT:    ; implicit-def: $sgpr38
; VI-NEXT:    ; implicit-def: $sgpr36
; VI-NEXT:    ; implicit-def: $sgpr34
; VI-NEXT:    ; implicit-def: $sgpr30
; VI-NEXT:    ; implicit-def: $sgpr90
; VI-NEXT:    ; implicit-def: $sgpr88
; VI-NEXT:    ; implicit-def: $sgpr78
; VI-NEXT:    ; implicit-def: $sgpr76
; VI-NEXT:    ; implicit-def: $sgpr74
; VI-NEXT:    ; implicit-def: $sgpr72
; VI-NEXT:    ; implicit-def: $sgpr62
; VI-NEXT:    ; implicit-def: $sgpr60
; VI-NEXT:    ; implicit-def: $sgpr58
; VI-NEXT:    v_writelane_b32 v33, s71, 1
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    s_branch .LBB13_2
;
; GFX9-LABEL: bitcast_v32i32_to_v128i8_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v29, s30, 0
; GFX9-NEXT:    v_writelane_b32 v29, s31, 1
; GFX9-NEXT:    v_writelane_b32 v29, s34, 2
; GFX9-NEXT:    v_writelane_b32 v29, s35, 3
; GFX9-NEXT:    v_writelane_b32 v29, s36, 4
; GFX9-NEXT:    v_writelane_b32 v29, s37, 5
; GFX9-NEXT:    v_writelane_b32 v29, s38, 6
; GFX9-NEXT:    v_writelane_b32 v29, s39, 7
; GFX9-NEXT:    v_writelane_b32 v29, s48, 8
; GFX9-NEXT:    v_writelane_b32 v29, s49, 9
; GFX9-NEXT:    v_writelane_b32 v29, s50, 10
; GFX9-NEXT:    v_writelane_b32 v29, s51, 11
; GFX9-NEXT:    v_writelane_b32 v29, s52, 12
; GFX9-NEXT:    v_writelane_b32 v29, s53, 13
; GFX9-NEXT:    v_writelane_b32 v29, s54, 14
; GFX9-NEXT:    v_writelane_b32 v29, s55, 15
; GFX9-NEXT:    v_writelane_b32 v29, s64, 16
; GFX9-NEXT:    v_writelane_b32 v29, s65, 17
; GFX9-NEXT:    v_writelane_b32 v29, s66, 18
; GFX9-NEXT:    v_writelane_b32 v29, s67, 19
; GFX9-NEXT:    v_writelane_b32 v29, s68, 20
; GFX9-NEXT:    v_writelane_b32 v29, s69, 21
; GFX9-NEXT:    v_writelane_b32 v29, s70, 22
; GFX9-NEXT:    v_writelane_b32 v29, s71, 23
; GFX9-NEXT:    v_writelane_b32 v29, s80, 24
; GFX9-NEXT:    v_writelane_b32 v29, s81, 25
; GFX9-NEXT:    v_writelane_b32 v29, s82, 26
; GFX9-NEXT:    v_writelane_b32 v29, s83, 27
; GFX9-NEXT:    v_writelane_b32 v29, s84, 28
; GFX9-NEXT:    v_writelane_b32 v29, s85, 29
; GFX9-NEXT:    v_writelane_b32 v29, s86, 30
; GFX9-NEXT:    v_writelane_b32 v29, s87, 31
; GFX9-NEXT:    v_writelane_b32 v29, s96, 32
; GFX9-NEXT:    v_writelane_b32 v29, s97, 33
; GFX9-NEXT:    v_readfirstlane_b32 s44, v19
; GFX9-NEXT:    v_writelane_b32 v29, s98, 34
; GFX9-NEXT:    v_readfirstlane_b32 s5, v18
; GFX9-NEXT:    v_readfirstlane_b32 s4, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s6, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s8, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s10, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s12, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s14, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s40, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s42, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    s_cmp_lg_u32 s44, 0
; GFX9-NEXT:    v_readfirstlane_b32 s44, v1
; GFX9-NEXT:    v_writelane_b32 v29, s99, 35
; GFX9-NEXT:    ; implicit-def: $vgpr30 : SGPR spill to VGPR lane
; GFX9-NEXT:    s_cbranch_scc0 .LBB13_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 0
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s6, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s6, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s8, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s8, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s11, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s11, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s11, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s10, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s10, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s13, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s13, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s13, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s12, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s12, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s15, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s15, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s15, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s14, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s14, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s41, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s41, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s41, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s40, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s40, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s43, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s43, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s43, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s42, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s42, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s45, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s45, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s45, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s44, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s44, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s26, 16
; GFX9-NEXT:    s_lshr_b32 s85, s29, 8
; GFX9-NEXT:    s_lshr_b32 s96, s28, 8
; GFX9-NEXT:    s_lshr_b32 s71, s27, 24
; GFX9-NEXT:    s_lshr_b32 s51, s27, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 49
; GFX9-NEXT:    s_lshr_b32 s67, s26, 8
; GFX9-NEXT:    s_lshr_b32 s82, s25, 24
; GFX9-NEXT:    s_lshr_b32 s84, s25, 16
; GFX9-NEXT:    s_lshr_b32 s65, s25, 8
; GFX9-NEXT:    s_lshr_b32 s86, s24, 16
; GFX9-NEXT:    s_lshr_b32 s55, s24, 8
; GFX9-NEXT:    s_lshr_b32 s98, s23, 24
; GFX9-NEXT:    s_lshr_b32 s87, s23, 16
; GFX9-NEXT:    s_lshr_b32 s66, s23, 8
; GFX9-NEXT:    s_lshr_b32 s97, s22, 16
; GFX9-NEXT:    s_lshr_b32 s64, s22, 8
; GFX9-NEXT:    s_lshr_b32 s49, s21, 24
; GFX9-NEXT:    s_lshr_b32 s99, s21, 16
; GFX9-NEXT:    s_lshr_b32 s80, s21, 8
; GFX9-NEXT:    s_lshr_b32 s39, s20, 16
; GFX9-NEXT:    s_lshr_b32 s70, s20, 8
; GFX9-NEXT:    s_lshr_b32 s48, s19, 24
; GFX9-NEXT:    s_lshr_b32 s38, s19, 16
; GFX9-NEXT:    s_lshr_b32 s68, s19, 8
; GFX9-NEXT:    s_lshr_b32 s54, s18, 16
; GFX9-NEXT:    s_lshr_b32 s69, s18, 8
; GFX9-NEXT:    s_lshr_b32 s52, s17, 24
; GFX9-NEXT:    s_lshr_b32 s50, s17, 16
; GFX9-NEXT:    s_lshr_b32 s83, s17, 8
; GFX9-NEXT:    s_lshr_b32 s53, s16, 16
; GFX9-NEXT:    s_lshr_b32 s81, s16, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[16:17], 24
; GFX9-NEXT:    s_cbranch_execnz .LBB13_3
; GFX9-NEXT:  .LBB13_2: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s5, s5, 3
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 0
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    s_add_i32 s4, s4, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s6, 16
; GFX9-NEXT:    s_add_i32 s9, s9, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s6, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s8, 16
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s8, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s11, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s11, 16
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s11, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s10, 16
; GFX9-NEXT:    s_add_i32 s13, s13, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s10, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s13, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s13, 16
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s13, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s12, 16
; GFX9-NEXT:    s_add_i32 s15, s15, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s12, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s15, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s15, 16
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s15, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s14, 16
; GFX9-NEXT:    s_add_i32 s41, s41, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s14, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s41, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s41, 16
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s41, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s40, 16
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s40, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s43, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s43, 16
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s43, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s42, 16
; GFX9-NEXT:    s_add_i32 s45, s45, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s42, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s45, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s45, 16
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s45, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s44, 16
; GFX9-NEXT:    s_add_i32 s29, s29, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s44, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    s_add_i32 s27, s27, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    s_add_i32 s17, s17, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    s_add_i32 s19, s19, 3
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    s_add_i32 s21, s21, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s23, s23, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s25, s25, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    v_writelane_b32 v30, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s26, 16
; GFX9-NEXT:    s_lshr_b32 s85, s29, 8
; GFX9-NEXT:    s_lshr_b32 s96, s28, 8
; GFX9-NEXT:    s_lshr_b32 s71, s27, 24
; GFX9-NEXT:    s_lshr_b32 s51, s27, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 49
; GFX9-NEXT:    s_lshr_b32 s67, s26, 8
; GFX9-NEXT:    s_lshr_b32 s82, s25, 24
; GFX9-NEXT:    s_lshr_b32 s84, s25, 16
; GFX9-NEXT:    s_lshr_b32 s65, s25, 8
; GFX9-NEXT:    s_lshr_b32 s86, s24, 16
; GFX9-NEXT:    s_lshr_b32 s55, s24, 8
; GFX9-NEXT:    s_lshr_b32 s98, s23, 24
; GFX9-NEXT:    s_lshr_b32 s87, s23, 16
; GFX9-NEXT:    s_lshr_b32 s66, s23, 8
; GFX9-NEXT:    s_lshr_b32 s97, s22, 16
; GFX9-NEXT:    s_lshr_b32 s64, s22, 8
; GFX9-NEXT:    s_lshr_b32 s49, s21, 24
; GFX9-NEXT:    s_lshr_b32 s99, s21, 16
; GFX9-NEXT:    s_lshr_b32 s80, s21, 8
; GFX9-NEXT:    s_lshr_b32 s39, s20, 16
; GFX9-NEXT:    s_lshr_b32 s70, s20, 8
; GFX9-NEXT:    s_lshr_b32 s48, s19, 24
; GFX9-NEXT:    s_lshr_b32 s38, s19, 16
; GFX9-NEXT:    s_lshr_b32 s68, s19, 8
; GFX9-NEXT:    s_lshr_b32 s54, s18, 16
; GFX9-NEXT:    s_lshr_b32 s69, s18, 8
; GFX9-NEXT:    s_lshr_b32 s52, s17, 24
; GFX9-NEXT:    s_lshr_b32 s50, s17, 16
; GFX9-NEXT:    s_lshr_b32 s83, s17, 8
; GFX9-NEXT:    s_lshr_b32 s53, s16, 16
; GFX9-NEXT:    s_lshr_b32 s81, s16, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[16:17], 24
; GFX9-NEXT:  .LBB13_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v6, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v2, s81
; GFX9-NEXT:    v_perm_b32 v6, s53, v6, v1
; GFX9-NEXT:    v_perm_b32 v3, s16, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v23, s52
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v4, s83
; GFX9-NEXT:    v_perm_b32 v23, s50, v23, v1
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v6
; GFX9-NEXT:    v_perm_b32 v2, s17, v4, v1
; GFX9-NEXT:    v_mov_b32_e32 v10, s34
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v23
; GFX9-NEXT:    v_mov_b32_e32 v4, s69
; GFX9-NEXT:    v_perm_b32 v10, s54, v10, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_perm_b32 v5, s18, v4, v1
; GFX9-NEXT:    v_mov_b32_e32 v24, s48
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v7, s68
; GFX9-NEXT:    v_perm_b32 v24, s38, v24, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v5, v2
; GFX9-NEXT:    v_perm_b32 v4, s19, v7, v1
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v24
; GFX9-NEXT:    v_mov_b32_e32 v7, s70
; GFX9-NEXT:    v_perm_b32 v14, s39, v14, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX9-NEXT:    v_perm_b32 v8, s20, v7, v1
; GFX9-NEXT:    v_mov_b32_e32 v25, s49
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v14
; GFX9-NEXT:    v_mov_b32_e32 v9, s80
; GFX9-NEXT:    v_perm_b32 v25, s99, v25, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v8, v2
; GFX9-NEXT:    v_perm_b32 v7, s21, v9, v1
; GFX9-NEXT:    v_mov_b32_e32 v15, s94
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v25
; GFX9-NEXT:    v_mov_b32_e32 v9, s64
; GFX9-NEXT:    v_perm_b32 v15, s97, v15, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v7, v2
; GFX9-NEXT:    v_perm_b32 v11, s22, v9, v1
; GFX9-NEXT:    v_mov_b32_e32 v26, s98
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v15
; GFX9-NEXT:    v_mov_b32_e32 v12, s66
; GFX9-NEXT:    v_perm_b32 v26, s87, v26, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v11, v2
; GFX9-NEXT:    v_perm_b32 v9, s23, v12, v1
; GFX9-NEXT:    v_mov_b32_e32 v16, s92
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v26
; GFX9-NEXT:    v_mov_b32_e32 v12, s55
; GFX9-NEXT:    v_perm_b32 v16, s86, v16, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v9, v2
; GFX9-NEXT:    v_perm_b32 v13, s24, v12, v1
; GFX9-NEXT:    v_mov_b32_e32 v27, s82
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v16
; GFX9-NEXT:    v_mov_b32_e32 v17, s65
; GFX9-NEXT:    v_perm_b32 v27, s84, v27, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v13, v2
; GFX9-NEXT:    v_perm_b32 v12, s25, v17, v1
; GFX9-NEXT:    v_mov_b32_e32 v17, s90
; GFX9-NEXT:    v_readlane_b32 s16, v30, 49
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
; GFX9-NEXT:    v_mov_b32_e32 v18, s67
; GFX9-NEXT:    v_perm_b32 v17, s16, v17, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v12, v2
; GFX9-NEXT:    v_perm_b32 v18, s26, v18, v1
; GFX9-NEXT:    v_mov_b32_e32 v28, s71
; GFX9-NEXT:    v_readlane_b32 s16, v30, 48
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
; GFX9-NEXT:    v_mov_b32_e32 v19, s51
; GFX9-NEXT:    v_perm_b32 v28, s16, v28, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v18, v2
; GFX9-NEXT:    v_perm_b32 v19, s27, v19, v1
; GFX9-NEXT:    v_mov_b32_e32 v20, s88
; GFX9-NEXT:    v_readlane_b32 s16, v30, 47
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
; GFX9-NEXT:    v_mov_b32_e32 v21, s96
; GFX9-NEXT:    v_perm_b32 v3, s16, v20, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v19, v2
; GFX9-NEXT:    v_perm_b32 v21, s28, v21, v1
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v21, v2
; GFX9-NEXT:    v_readlane_b32 s16, v30, 45
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 46
; GFX9-NEXT:    v_mov_b32_e32 v22, s85
; GFX9-NEXT:    v_perm_b32 v2, s16, v2, v1
; GFX9-NEXT:    v_perm_b32 v22, s29, v22, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v2, v22, v2
; GFX9-NEXT:    v_readlane_b32 s16, v30, 44
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_mov_b32_e32 v3, s78
; GFX9-NEXT:    v_readlane_b32 s16, v30, 43
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s44, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 42
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 40
; GFX9-NEXT:    v_mov_b32_e32 v3, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 41
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s45, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 39
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_mov_b32_e32 v3, s76
; GFX9-NEXT:    v_readlane_b32 s16, v30, 38
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s42, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 37
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 35
; GFX9-NEXT:    v_mov_b32_e32 v3, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 36
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s43, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 34
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_mov_b32_e32 v3, s74
; GFX9-NEXT:    v_readlane_b32 s16, v30, 33
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s40, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 32
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 30
; GFX9-NEXT:    v_mov_b32_e32 v3, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 31
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s41, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 29
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_perm_b32 v2, s14, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s72
; GFX9-NEXT:    v_readlane_b32 s14, v30, 28
; GFX9-NEXT:    v_perm_b32 v3, s14, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s14, v30, 27
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    v_mov_b32_e32 v2, s14
; GFX9-NEXT:    v_readlane_b32 s14, v30, 25
; GFX9-NEXT:    v_mov_b32_e32 v3, s14
; GFX9-NEXT:    v_readlane_b32 s14, v30, 26
; GFX9-NEXT:    v_perm_b32 v3, s14, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s15, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s14, v30, 24
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    v_mov_b32_e32 v2, s14
; GFX9-NEXT:    v_perm_b32 v2, s12, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s62
; GFX9-NEXT:    v_readlane_b32 s12, v30, 23
; GFX9-NEXT:    v_perm_b32 v3, s12, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s12, v30, 22
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    v_mov_b32_e32 v2, s12
; GFX9-NEXT:    v_readlane_b32 s12, v30, 20
; GFX9-NEXT:    v_mov_b32_e32 v3, s12
; GFX9-NEXT:    v_readlane_b32 s12, v30, 21
; GFX9-NEXT:    v_perm_b32 v3, s12, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s13, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s12, v30, 19
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    v_mov_b32_e32 v2, s12
; GFX9-NEXT:    v_perm_b32 v2, s10, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s60
; GFX9-NEXT:    v_readlane_b32 s10, v30, 18
; GFX9-NEXT:    v_perm_b32 v3, s10, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s10, v30, 17
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    v_mov_b32_e32 v2, s10
; GFX9-NEXT:    v_readlane_b32 s10, v30, 15
; GFX9-NEXT:    v_mov_b32_e32 v3, s10
; GFX9-NEXT:    v_readlane_b32 s10, v30, 16
; GFX9-NEXT:    v_perm_b32 v3, s10, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s11, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s10, v30, 14
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    v_mov_b32_e32 v2, s10
; GFX9-NEXT:    v_perm_b32 v2, s8, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s58
; GFX9-NEXT:    v_readlane_b32 s8, v30, 13
; GFX9-NEXT:    v_perm_b32 v3, s8, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s8, v30, 12
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    v_mov_b32_e32 v2, s8
; GFX9-NEXT:    v_readlane_b32 s8, v30, 10
; GFX9-NEXT:    v_mov_b32_e32 v3, s8
; GFX9-NEXT:    v_readlane_b32 s8, v30, 11
; GFX9-NEXT:    v_perm_b32 v3, s8, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s9, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s8, v30, 9
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    v_mov_b32_e32 v2, s8
; GFX9-NEXT:    v_perm_b32 v2, s6, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s56
; GFX9-NEXT:    v_readlane_b32 s6, v30, 8
; GFX9-NEXT:    v_perm_b32 v3, s6, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s6, v30, 7
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    v_mov_b32_e32 v2, s6
; GFX9-NEXT:    v_readlane_b32 s6, v30, 5
; GFX9-NEXT:    v_mov_b32_e32 v3, s6
; GFX9-NEXT:    v_readlane_b32 s6, v30, 6
; GFX9-NEXT:    v_perm_b32 v3, s6, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s7, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s6, v30, 4
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    v_mov_b32_e32 v2, s6
; GFX9-NEXT:    v_perm_b32 v2, s4, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s46
; GFX9-NEXT:    v_readlane_b32 s4, v30, 3
; GFX9-NEXT:    v_perm_b32 v3, s4, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s4, v30, 2
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    v_mov_b32_e32 v2, s4
; GFX9-NEXT:    v_readlane_b32 s4, v30, 0
; GFX9-NEXT:    v_mov_b32_e32 v3, s4
; GFX9-NEXT:    v_readlane_b32 s4, v30, 1
; GFX9-NEXT:    v_perm_b32 v2, s5, v2, v1
; GFX9-NEXT:    v_perm_b32 v1, s4, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    v_readlane_b32 s99, v29, 35
; GFX9-NEXT:    v_readlane_b32 s98, v29, 34
; GFX9-NEXT:    v_readlane_b32 s97, v29, 33
; GFX9-NEXT:    v_readlane_b32 s96, v29, 32
; GFX9-NEXT:    v_readlane_b32 s87, v29, 31
; GFX9-NEXT:    v_readlane_b32 s86, v29, 30
; GFX9-NEXT:    v_readlane_b32 s85, v29, 29
; GFX9-NEXT:    v_readlane_b32 s84, v29, 28
; GFX9-NEXT:    v_readlane_b32 s83, v29, 27
; GFX9-NEXT:    v_readlane_b32 s82, v29, 26
; GFX9-NEXT:    v_readlane_b32 s81, v29, 25
; GFX9-NEXT:    v_readlane_b32 s80, v29, 24
; GFX9-NEXT:    v_readlane_b32 s71, v29, 23
; GFX9-NEXT:    v_readlane_b32 s70, v29, 22
; GFX9-NEXT:    v_readlane_b32 s69, v29, 21
; GFX9-NEXT:    v_readlane_b32 s68, v29, 20
; GFX9-NEXT:    v_readlane_b32 s67, v29, 19
; GFX9-NEXT:    v_readlane_b32 s66, v29, 18
; GFX9-NEXT:    v_readlane_b32 s65, v29, 17
; GFX9-NEXT:    v_readlane_b32 s64, v29, 16
; GFX9-NEXT:    v_readlane_b32 s55, v29, 15
; GFX9-NEXT:    v_readlane_b32 s54, v29, 14
; GFX9-NEXT:    v_readlane_b32 s53, v29, 13
; GFX9-NEXT:    v_readlane_b32 s52, v29, 12
; GFX9-NEXT:    v_readlane_b32 s51, v29, 11
; GFX9-NEXT:    v_readlane_b32 s50, v29, 10
; GFX9-NEXT:    v_readlane_b32 s49, v29, 9
; GFX9-NEXT:    v_readlane_b32 s48, v29, 8
; GFX9-NEXT:    v_readlane_b32 s39, v29, 7
; GFX9-NEXT:    v_readlane_b32 s38, v29, 6
; GFX9-NEXT:    v_readlane_b32 s37, v29, 5
; GFX9-NEXT:    v_readlane_b32 s36, v29, 4
; GFX9-NEXT:    v_readlane_b32 s35, v29, 3
; GFX9-NEXT:    v_readlane_b32 s34, v29, 2
; GFX9-NEXT:    v_readlane_b32 s31, v29, 1
; GFX9-NEXT:    v_readlane_b32 s30, v29, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB13_4:
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr81
; GFX9-NEXT:    ; implicit-def: $sgpr53
; GFX9-NEXT:    ; implicit-def: $sgpr83
; GFX9-NEXT:    ; implicit-def: $sgpr50
; GFX9-NEXT:    ; implicit-def: $sgpr52
; GFX9-NEXT:    ; implicit-def: $sgpr69
; GFX9-NEXT:    ; implicit-def: $sgpr54
; GFX9-NEXT:    ; implicit-def: $sgpr68
; GFX9-NEXT:    ; implicit-def: $sgpr38
; GFX9-NEXT:    ; implicit-def: $sgpr48
; GFX9-NEXT:    ; implicit-def: $sgpr70
; GFX9-NEXT:    ; implicit-def: $sgpr39
; GFX9-NEXT:    ; implicit-def: $sgpr80
; GFX9-NEXT:    ; implicit-def: $sgpr99
; GFX9-NEXT:    ; implicit-def: $sgpr49
; GFX9-NEXT:    ; implicit-def: $sgpr64
; GFX9-NEXT:    ; implicit-def: $sgpr97
; GFX9-NEXT:    ; implicit-def: $sgpr66
; GFX9-NEXT:    ; implicit-def: $sgpr87
; GFX9-NEXT:    ; implicit-def: $sgpr98
; GFX9-NEXT:    ; implicit-def: $sgpr55
; GFX9-NEXT:    ; implicit-def: $sgpr86
; GFX9-NEXT:    ; implicit-def: $sgpr65
; GFX9-NEXT:    ; implicit-def: $sgpr84
; GFX9-NEXT:    ; implicit-def: $sgpr82
; GFX9-NEXT:    ; implicit-def: $sgpr67
; GFX9-NEXT:    ; implicit-def: $sgpr51
; GFX9-NEXT:    ; implicit-def: $sgpr71
; GFX9-NEXT:    ; implicit-def: $sgpr96
; GFX9-NEXT:    ; implicit-def: $sgpr85
; GFX9-NEXT:    ; implicit-def: $sgpr36
; GFX9-NEXT:    ; implicit-def: $sgpr34
; GFX9-NEXT:    ; implicit-def: $sgpr30
; GFX9-NEXT:    ; implicit-def: $sgpr94
; GFX9-NEXT:    ; implicit-def: $sgpr92
; GFX9-NEXT:    ; implicit-def: $sgpr90
; GFX9-NEXT:    ; implicit-def: $sgpr88
; GFX9-NEXT:    ; implicit-def: $sgpr78
; GFX9-NEXT:    ; implicit-def: $sgpr76
; GFX9-NEXT:    ; implicit-def: $sgpr74
; GFX9-NEXT:    ; implicit-def: $sgpr72
; GFX9-NEXT:    ; implicit-def: $sgpr62
; GFX9-NEXT:    ; implicit-def: $sgpr60
; GFX9-NEXT:    ; implicit-def: $sgpr58
; GFX9-NEXT:    ; implicit-def: $sgpr56
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    s_branch .LBB13_2
;
; GFX11-LABEL: bitcast_v32i32_to_v128i8_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v24, s32
; GFX11-NEXT:    scratch_store_b32 off, v25, s32 offset:4
; GFX11-NEXT:    scratch_store_b32 off, v26, s32 offset:8
; GFX11-NEXT:    scratch_store_b32 off, v27, s32 offset:12
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v24, s30, 0
; GFX11-NEXT:    v_writelane_b32 v25, s96, 0
; GFX11-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_writelane_b32 v24, s31, 1
; GFX11-NEXT:    v_writelane_b32 v25, s97, 1
; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
; GFX11-NEXT:    v_writelane_b32 v24, s34, 2
; GFX11-NEXT:    v_writelane_b32 v25, s98, 2
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_writelane_b32 v24, s35, 3
; GFX11-NEXT:    v_writelane_b32 v25, s99, 3
; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
; GFX11-NEXT:    v_writelane_b32 v24, s36, 4
; GFX11-NEXT:    v_writelane_b32 v25, s100, 4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_writelane_b32 v24, s37, 5
; GFX11-NEXT:    v_writelane_b32 v25, s101, 5
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-NEXT:    ; implicit-def: $vgpr27 : SGPR spill to VGPR lane
; GFX11-NEXT:    ; implicit-def: $vgpr26 : SGPR spill to VGPR lane
; GFX11-NEXT:    v_writelane_b32 v24, s38, 6
; GFX11-NEXT:    v_writelane_b32 v25, s102, 6
; GFX11-NEXT:    v_writelane_b32 v24, s39, 7
; GFX11-NEXT:    v_writelane_b32 v25, s103, 7
; GFX11-NEXT:    v_writelane_b32 v24, s48, 8
; GFX11-NEXT:    v_writelane_b32 v25, s104, 8
; GFX11-NEXT:    v_writelane_b32 v24, s49, 9
; GFX11-NEXT:    v_writelane_b32 v24, s50, 10
; GFX11-NEXT:    v_writelane_b32 v24, s51, 11
; GFX11-NEXT:    v_writelane_b32 v24, s52, 12
; GFX11-NEXT:    v_writelane_b32 v24, s53, 13
; GFX11-NEXT:    v_writelane_b32 v24, s54, 14
; GFX11-NEXT:    v_writelane_b32 v24, s55, 15
; GFX11-NEXT:    v_writelane_b32 v24, s64, 16
; GFX11-NEXT:    v_writelane_b32 v24, s65, 17
; GFX11-NEXT:    v_writelane_b32 v24, s66, 18
; GFX11-NEXT:    v_writelane_b32 v24, s67, 19
; GFX11-NEXT:    v_writelane_b32 v24, s68, 20
; GFX11-NEXT:    v_writelane_b32 v24, s69, 21
; GFX11-NEXT:    v_writelane_b32 v24, s70, 22
; GFX11-NEXT:    v_writelane_b32 v24, s71, 23
; GFX11-NEXT:    v_writelane_b32 v24, s80, 24
; GFX11-NEXT:    v_writelane_b32 v24, s81, 25
; GFX11-NEXT:    v_writelane_b32 v24, s82, 26
; GFX11-NEXT:    v_writelane_b32 v24, s83, 27
; GFX11-NEXT:    v_writelane_b32 v24, s84, 28
; GFX11-NEXT:    v_writelane_b32 v24, s85, 29
; GFX11-NEXT:    v_writelane_b32 v24, s86, 30
; GFX11-NEXT:    v_writelane_b32 v24, s87, 31
; GFX11-NEXT:    s_cbranch_scc0 .LBB13_4
; GFX11-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s27, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
; GFX11-NEXT:    s_lshr_b32 s34, s27, 16
; GFX11-NEXT:    s_lshr_b32 s35, s27, 8
; GFX11-NEXT:    s_lshr_b32 s37, s26, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
; GFX11-NEXT:    s_lshr_b32 s36, s26, 8
; GFX11-NEXT:    s_lshr_b32 s39, s25, 24
; GFX11-NEXT:    s_lshr_b32 s48, s25, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s4, 16
; GFX11-NEXT:    s_lshr_b32 s38, s25, 8
; GFX11-NEXT:    s_lshr_b32 s54, s24, 16
; GFX11-NEXT:    s_lshr_b32 s49, s24, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s4, 8
; GFX11-NEXT:    s_lshr_b32 s50, s23, 24
; GFX11-NEXT:    s_lshr_b32 s52, s23, 16
; GFX11-NEXT:    s_lshr_b32 s51, s23, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
; GFX11-NEXT:    s_lshr_b32 s80, s22, 16
; GFX11-NEXT:    s_lshr_b32 s53, s22, 8
; GFX11-NEXT:    s_lshr_b32 s55, s21, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
; GFX11-NEXT:    s_lshr_b32 s64, s21, 16
; GFX11-NEXT:    s_lshr_b32 s69, s21, 8
; GFX11-NEXT:    s_lshr_b32 s84, s20, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
; GFX11-NEXT:    s_lshr_b32 s82, s20, 8
; GFX11-NEXT:    s_lshr_b32 s65, s19, 24
; GFX11-NEXT:    s_lshr_b32 s66, s19, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s6, 16
; GFX11-NEXT:    s_lshr_b32 s85, s19, 8
; GFX11-NEXT:    s_lshr_b32 s86, s18, 16
; GFX11-NEXT:    s_lshr_b32 s87, s18, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s6, 8
; GFX11-NEXT:    s_lshr_b32 s67, s17, 24
; GFX11-NEXT:    s_lshr_b32 s68, s17, 16
; GFX11-NEXT:    s_lshr_b32 s96, s17, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
; GFX11-NEXT:    s_lshr_b32 s97, s16, 16
; GFX11-NEXT:    s_lshr_b32 s98, s16, 8
; GFX11-NEXT:    s_lshr_b32 s70, s3, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
; GFX11-NEXT:    s_lshr_b32 s71, s3, 16
; GFX11-NEXT:    s_lshr_b32 s99, s3, 8
; GFX11-NEXT:    s_lshr_b32 s100, s2, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
; GFX11-NEXT:    s_lshr_b32 s101, s2, 8
; GFX11-NEXT:    s_lshr_b32 s81, s1, 24
; GFX11-NEXT:    s_lshr_b32 s83, s1, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s8, 16
; GFX11-NEXT:    s_lshr_b32 s102, s1, 8
; GFX11-NEXT:    s_lshr_b32 s103, s0, 16
; GFX11-NEXT:    s_lshr_b32 s104, s0, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s8, 8
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[6:7], 24
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[10:11], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s11, 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[24:25], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s11, 8
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[18:19], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s10, 16
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[16:17], 24
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[2:3], 24
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s10, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s13, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s12, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s12, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s15, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s15, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s15, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s14, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s14, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s41, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s41, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s41, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s40, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s40, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 7
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[4:5], 24
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-NEXT:    s_cbranch_vccnz .LBB13_3
; GFX11-NEXT:  .LBB13_2: ; %cmp.true
; GFX11-NEXT:    s_add_i32 s5, s5, 3
; GFX11-NEXT:    s_add_i32 s4, s4, 3
; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
; GFX11-NEXT:    s_add_i32 s7, s7, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
; GFX11-NEXT:    s_add_i32 s6, s6, 3
; GFX11-NEXT:    s_add_i32 s9, s9, 3
; GFX11-NEXT:    s_add_i32 s8, s8, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
; GFX11-NEXT:    s_add_i32 s11, s11, 3
; GFX11-NEXT:    s_add_i32 s10, s10, 3
; GFX11-NEXT:    s_add_i32 s13, s13, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s4, 16
; GFX11-NEXT:    s_add_i32 s12, s12, 3
; GFX11-NEXT:    s_add_i32 s15, s15, 3
; GFX11-NEXT:    s_add_i32 s14, s14, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s4, 8
; GFX11-NEXT:    s_add_i32 s41, s41, 3
; GFX11-NEXT:    s_add_i32 s40, s40, 3
; GFX11-NEXT:    s_add_i32 s29, s29, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
; GFX11-NEXT:    s_add_i32 s28, s28, 3
; GFX11-NEXT:    s_add_i32 s1, s1, 3
; GFX11-NEXT:    s_add_i32 s0, s0, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
; GFX11-NEXT:    s_add_i32 s3, s3, 3
; GFX11-NEXT:    s_add_i32 s2, s2, 3
; GFX11-NEXT:    s_add_i32 s17, s17, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
; GFX11-NEXT:    s_add_i32 s16, s16, 3
; GFX11-NEXT:    s_add_i32 s19, s19, 3
; GFX11-NEXT:    s_add_i32 s18, s18, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s6, 16
; GFX11-NEXT:    s_add_i32 s21, s21, 3
; GFX11-NEXT:    s_add_i32 s20, s20, 3
; GFX11-NEXT:    s_add_i32 s23, s23, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s6, 8
; GFX11-NEXT:    s_add_i32 s22, s22, 3
; GFX11-NEXT:    s_add_i32 s25, s25, 3
; GFX11-NEXT:    s_add_i32 s24, s24, 3
; GFX11-NEXT:    v_writelane_b32 v27, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
; GFX11-NEXT:    s_add_i32 s27, s27, 3
; GFX11-NEXT:    s_add_i32 s26, s26, 3
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s27, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
; GFX11-NEXT:    s_lshr_b32 s34, s27, 16
; GFX11-NEXT:    s_lshr_b32 s35, s27, 8
; GFX11-NEXT:    s_lshr_b32 s37, s26, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
; GFX11-NEXT:    s_lshr_b32 s36, s26, 8
; GFX11-NEXT:    s_lshr_b32 s39, s25, 24
; GFX11-NEXT:    s_lshr_b32 s48, s25, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s8, 16
; GFX11-NEXT:    s_lshr_b32 s38, s25, 8
; GFX11-NEXT:    s_lshr_b32 s54, s24, 16
; GFX11-NEXT:    s_lshr_b32 s49, s24, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s8, 8
; GFX11-NEXT:    s_lshr_b32 s50, s23, 24
; GFX11-NEXT:    s_lshr_b32 s52, s23, 16
; GFX11-NEXT:    s_lshr_b32 s51, s23, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s11, 24
; GFX11-NEXT:    s_lshr_b32 s80, s22, 16
; GFX11-NEXT:    s_lshr_b32 s53, s22, 8
; GFX11-NEXT:    s_lshr_b32 s55, s21, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
; GFX11-NEXT:    s_lshr_b32 s64, s21, 16
; GFX11-NEXT:    s_lshr_b32 s69, s21, 8
; GFX11-NEXT:    s_lshr_b32 s84, s20, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s11, 8
; GFX11-NEXT:    s_lshr_b32 s82, s20, 8
; GFX11-NEXT:    s_lshr_b32 s65, s19, 24
; GFX11-NEXT:    s_lshr_b32 s66, s19, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s10, 16
; GFX11-NEXT:    s_lshr_b32 s85, s19, 8
; GFX11-NEXT:    s_lshr_b32 s86, s18, 16
; GFX11-NEXT:    s_lshr_b32 s87, s18, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s10, 8
; GFX11-NEXT:    s_lshr_b32 s67, s17, 24
; GFX11-NEXT:    s_lshr_b32 s68, s17, 16
; GFX11-NEXT:    s_lshr_b32 s96, s17, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s13, 24
; GFX11-NEXT:    s_lshr_b32 s97, s16, 16
; GFX11-NEXT:    s_lshr_b32 s98, s16, 8
; GFX11-NEXT:    s_lshr_b32 s70, s3, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
; GFX11-NEXT:    s_lshr_b32 s71, s3, 16
; GFX11-NEXT:    s_lshr_b32 s99, s3, 8
; GFX11-NEXT:    s_lshr_b32 s100, s2, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
; GFX11-NEXT:    s_lshr_b32 s101, s2, 8
; GFX11-NEXT:    s_lshr_b32 s81, s1, 24
; GFX11-NEXT:    s_lshr_b32 s83, s1, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s12, 16
; GFX11-NEXT:    s_lshr_b32 s102, s1, 8
; GFX11-NEXT:    s_lshr_b32 s103, s0, 16
; GFX11-NEXT:    s_lshr_b32 s104, s0, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s12, 8
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[6:7], 24
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[10:11], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s15, 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s15, 16
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[24:25], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s15, 8
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[18:19], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s14, 16
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[16:17], 24
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[2:3], 24
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s14, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s41, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s41, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s41, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s40, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s40, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 7
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[4:5], 24
; GFX11-NEXT:  .LBB13_3: ; %end
; GFX11-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; GFX11-NEXT:    v_readlane_b32 s31, v24, 1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_perm_b32 v2, s103, s30, v1
; GFX11-NEXT:    v_readlane_b32 s103, v25, 7
; GFX11-NEXT:    v_readlane_b32 s30, v24, 0
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-NEXT:    v_perm_b32 v19, s83, s81, v1
; GFX11-NEXT:    v_perm_b32 v3, s0, s104, v1
; GFX11-NEXT:    v_perm_b32 v4, s1, s102, v1
; GFX11-NEXT:    v_perm_b32 v5, s100, s92, v1
; GFX11-NEXT:    v_perm_b32 v20, s71, s70, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-NEXT:    v_perm_b32 v6, s2, s101, v1
; GFX11-NEXT:    v_perm_b32 v7, s3, s99, v1
; GFX11-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-NEXT:    v_perm_b32 v8, s97, s90, v1
; GFX11-NEXT:    v_or_b32_e32 v3, v4, v19
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v20
; GFX11-NEXT:    v_perm_b32 v11, s86, s88, v1
; GFX11-NEXT:    v_perm_b32 v21, s68, s67, v1
; GFX11-NEXT:    v_perm_b32 v22, s66, s65, v1
; GFX11-NEXT:    v_or_b32_e32 v4, v6, v4
; GFX11-NEXT:    v_or_b32_e32 v5, v7, v5
; GFX11-NEXT:    v_perm_b32 v9, s16, s98, v1
; GFX11-NEXT:    v_perm_b32 v10, s17, s96, v1
; GFX11-NEXT:    v_perm_b32 v12, s18, s87, v1
; GFX11-NEXT:    v_perm_b32 v13, s19, s85, v1
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v8
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v21
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v11
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v22
; GFX11-NEXT:    v_perm_b32 v15, s80, s76, v1
; GFX11-NEXT:    v_perm_b32 v23, s64, s55, v1
; GFX11-NEXT:    v_or_b32_e32 v2, v9, v2
; GFX11-NEXT:    v_or_b32_e32 v3, v10, v3
; GFX11-NEXT:    v_or_b32_e32 v4, v12, v4
; GFX11-NEXT:    v_or_b32_e32 v5, v13, v5
; GFX11-NEXT:    v_perm_b32 v14, s84, s78, v1
; GFX11-NEXT:    v_perm_b32 v17, s21, s69, v1
; GFX11-NEXT:    v_perm_b32 v18, s54, s94, v1
; GFX11-NEXT:    v_perm_b32 v6, s22, s53, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v23
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v15
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:16
; GFX11-NEXT:    v_perm_b32 v5, s52, s50, v1
; GFX11-NEXT:    v_perm_b32 v16, s20, s82, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v14
; GFX11-NEXT:    v_or_b32_e32 v3, v17, v8
; GFX11-NEXT:    v_or_b32_e32 v4, v6, v9
; GFX11-NEXT:    v_perm_b32 v6, s23, s51, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-NEXT:    v_perm_b32 v8, s24, s49, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v18
; GFX11-NEXT:    v_perm_b32 v11, s37, s74, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 6
; GFX11-NEXT:    v_or_b32_e32 v2, v16, v7
; GFX11-NEXT:    v_perm_b32 v7, s48, s39, v1
; GFX11-NEXT:    v_or_b32_e32 v5, v6, v5
; GFX11-NEXT:    v_or_b32_e32 v6, v8, v9
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v11
; GFX11-NEXT:    v_perm_b32 v11, s0, s72, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 3
; GFX11-NEXT:    v_readlane_b32 s1, v26, 4
; GFX11-NEXT:    v_perm_b32 v10, s25, s38, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:32
; GFX11-NEXT:    v_perm_b32 v8, s26, s36, v1
; GFX11-NEXT:    v_perm_b32 v2, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 7
; GFX11-NEXT:    v_or_b32_e32 v7, v10, v7
; GFX11-NEXT:    v_perm_b32 v10, s34, vcc_hi, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v11
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-NEXT:    v_perm_b32 v4, s28, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 5
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v10
; GFX11-NEXT:    v_perm_b32 v9, s27, s35, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s29, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 1
; GFX11-NEXT:    v_readlane_b32 s1, v27, 31
; GFX11-NEXT:    v_or_b32_e32 v9, v9, v3
; GFX11-NEXT:    v_readlane_b32 s104, v25, 8
; GFX11-NEXT:    v_or_b32_e32 v3, v10, v11
; GFX11-NEXT:    v_perm_b32 v12, s0, s62, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 2
; GFX11-NEXT:    scratch_store_b128 v0, v[6:9], off offset:48
; GFX11-NEXT:    v_readlane_b32 s102, v25, 6
; GFX11-NEXT:    v_readlane_b32 s101, v25, 5
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v12
; GFX11-NEXT:    v_perm_b32 v4, s40, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 30
; GFX11-NEXT:    v_readlane_b32 s100, v25, 4
; GFX11-NEXT:    v_readlane_b32 s99, v25, 3
; GFX11-NEXT:    v_readlane_b32 s98, v25, 2
; GFX11-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 0
; GFX11-NEXT:    v_readlane_b32 s1, v27, 26
; GFX11-NEXT:    v_readlane_b32 s97, v25, 1
; GFX11-NEXT:    v_readlane_b32 s96, v25, 0
; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v10
; GFX11-NEXT:    v_perm_b32 v5, s41, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 28
; GFX11-NEXT:    v_readlane_b32 s87, v24, 31
; GFX11-NEXT:    v_readlane_b32 s86, v24, 30
; GFX11-NEXT:    v_readlane_b32 s85, v24, 29
; GFX11-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX11-NEXT:    v_perm_b32 v11, s0, s60, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 25
; GFX11-NEXT:    v_readlane_b32 s84, v24, 28
; GFX11-NEXT:    v_readlane_b32 s83, v24, 27
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:64
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v11
; GFX11-NEXT:    v_perm_b32 v6, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 29
; GFX11-NEXT:    v_readlane_b32 s1, v27, 21
; GFX11-NEXT:    v_readlane_b32 s82, v24, 26
; GFX11-NEXT:    v_readlane_b32 s81, v24, 25
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v6
; GFX11-NEXT:    v_perm_b32 v8, s14, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 27
; GFX11-NEXT:    v_readlane_b32 s80, v24, 24
; GFX11-NEXT:    v_readlane_b32 s71, v24, 23
; GFX11-NEXT:    v_readlane_b32 s70, v24, 22
; GFX11-NEXT:    v_or_b32_e32 v6, v8, v9
; GFX11-NEXT:    v_perm_b32 v10, s15, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 23
; GFX11-NEXT:    v_readlane_b32 s69, v24, 21
; GFX11-NEXT:    v_readlane_b32 s68, v24, 20
; GFX11-NEXT:    v_readlane_b32 s67, v24, 19
; GFX11-NEXT:    v_or_b32_e32 v7, v10, v11
; GFX11-NEXT:    v_perm_b32 v12, s0, s58, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 24
; GFX11-NEXT:    v_readlane_b32 s66, v24, 18
; GFX11-NEXT:    v_readlane_b32 s65, v24, 17
; GFX11-NEXT:    v_readlane_b32 s64, v24, 16
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v12
; GFX11-NEXT:    v_perm_b32 v8, s12, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 20
; GFX11-NEXT:    v_readlane_b32 s55, v24, 15
; GFX11-NEXT:    v_readlane_b32 s54, v24, 14
; GFX11-NEXT:    v_readlane_b32 s53, v24, 13
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 22
; GFX11-NEXT:    v_readlane_b32 s1, v27, 16
; GFX11-NEXT:    v_readlane_b32 s52, v24, 12
; GFX11-NEXT:    v_readlane_b32 s51, v24, 11
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v10
; GFX11-NEXT:    v_perm_b32 v9, s13, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 18
; GFX11-NEXT:    v_readlane_b32 s50, v24, 10
; GFX11-NEXT:    v_readlane_b32 s49, v24, 9
; GFX11-NEXT:    v_readlane_b32 s48, v24, 8
; GFX11-NEXT:    v_or_b32_e32 v9, v9, v3
; GFX11-NEXT:    v_perm_b32 v11, s0, s56, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 15
; GFX11-NEXT:    v_readlane_b32 s39, v24, 7
; GFX11-NEXT:    v_readlane_b32 s38, v24, 6
; GFX11-NEXT:    v_readlane_b32 s37, v24, 5
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v11
; GFX11-NEXT:    v_perm_b32 v2, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 19
; GFX11-NEXT:    v_readlane_b32 s1, v27, 11
; GFX11-NEXT:    v_readlane_b32 s36, v24, 4
; GFX11-NEXT:    v_readlane_b32 s35, v24, 3
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX11-NEXT:    v_perm_b32 v4, s10, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 17
; GFX11-NEXT:    v_readlane_b32 s34, v24, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s11, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 13
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v3, v10, v11
; GFX11-NEXT:    v_perm_b32 v12, s0, s46, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 14
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v12
; GFX11-NEXT:    v_perm_b32 v4, s8, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 10
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 12
; GFX11-NEXT:    v_readlane_b32 s1, v27, 6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_perm_b32 v5, s9, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 8
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v5, v5, v10
; GFX11-NEXT:    v_perm_b32 v11, s0, s44, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 9
; GFX11-NEXT:    v_readlane_b32 s1, v27, 1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v10
; GFX11-NEXT:    v_perm_b32 v12, s6, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 7
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v10, v12, v11
; GFX11-NEXT:    v_perm_b32 v14, s7, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v11, v14, v13
; GFX11-NEXT:    v_perm_b32 v12, s0, s42, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-NEXT:    v_perm_b32 v13, s4, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX11-NEXT:    v_perm_b32 v13, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-NEXT:    v_perm_b32 v1, s5, s0, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_or_b32_e32 v13, v1, v13
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    scratch_store_b128 v0, v[6:9], off offset:80
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:96
; GFX11-NEXT:    scratch_store_b128 v0, v[10:13], off offset:112
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v24, off, s32
; GFX11-NEXT:    scratch_load_b32 v25, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v26, off, s32 offset:8
; GFX11-NEXT:    scratch_load_b32 v27, off, s32 offset:12
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB13_4:
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr104
; GFX11-NEXT:    ; implicit-def: $sgpr103
; GFX11-NEXT:    ; implicit-def: $sgpr30
; GFX11-NEXT:    ; implicit-def: $sgpr102
; GFX11-NEXT:    ; implicit-def: $sgpr83
; GFX11-NEXT:    ; implicit-def: $sgpr81
; GFX11-NEXT:    ; implicit-def: $sgpr101
; GFX11-NEXT:    ; implicit-def: $sgpr100
; GFX11-NEXT:    ; implicit-def: $sgpr92
; GFX11-NEXT:    ; implicit-def: $sgpr99
; GFX11-NEXT:    ; implicit-def: $sgpr71
; GFX11-NEXT:    ; implicit-def: $sgpr70
; GFX11-NEXT:    ; implicit-def: $sgpr98
; GFX11-NEXT:    ; implicit-def: $sgpr97
; GFX11-NEXT:    ; implicit-def: $sgpr90
; GFX11-NEXT:    ; implicit-def: $sgpr96
; GFX11-NEXT:    ; implicit-def: $sgpr68
; GFX11-NEXT:    ; implicit-def: $sgpr67
; GFX11-NEXT:    ; implicit-def: $sgpr87
; GFX11-NEXT:    ; implicit-def: $sgpr86
; GFX11-NEXT:    ; implicit-def: $sgpr88
; GFX11-NEXT:    ; implicit-def: $sgpr85
; GFX11-NEXT:    ; implicit-def: $sgpr66
; GFX11-NEXT:    ; implicit-def: $sgpr65
; GFX11-NEXT:    ; implicit-def: $sgpr82
; GFX11-NEXT:    ; implicit-def: $sgpr84
; GFX11-NEXT:    ; implicit-def: $sgpr78
; GFX11-NEXT:    ; implicit-def: $sgpr69
; GFX11-NEXT:    ; implicit-def: $sgpr64
; GFX11-NEXT:    ; implicit-def: $sgpr55
; GFX11-NEXT:    ; implicit-def: $sgpr53
; GFX11-NEXT:    ; implicit-def: $sgpr80
; GFX11-NEXT:    ; implicit-def: $sgpr76
; GFX11-NEXT:    ; implicit-def: $sgpr51
; GFX11-NEXT:    ; implicit-def: $sgpr52
; GFX11-NEXT:    ; implicit-def: $sgpr50
; GFX11-NEXT:    ; implicit-def: $sgpr49
; GFX11-NEXT:    ; implicit-def: $sgpr54
; GFX11-NEXT:    ; implicit-def: $sgpr38
; GFX11-NEXT:    ; implicit-def: $sgpr48
; GFX11-NEXT:    ; implicit-def: $sgpr39
; GFX11-NEXT:    ; implicit-def: $sgpr36
; GFX11-NEXT:    ; implicit-def: $sgpr37
; GFX11-NEXT:    ; implicit-def: $sgpr35
; GFX11-NEXT:    ; implicit-def: $sgpr34
; GFX11-NEXT:    ; implicit-def: $vcc_hi
; GFX11-NEXT:    ; implicit-def: $sgpr94
; GFX11-NEXT:    ; implicit-def: $sgpr74
; GFX11-NEXT:    ; implicit-def: $sgpr72
; GFX11-NEXT:    ; implicit-def: $sgpr62
; GFX11-NEXT:    ; implicit-def: $sgpr60
; GFX11-NEXT:    ; implicit-def: $sgpr58
; GFX11-NEXT:    ; implicit-def: $sgpr56
; GFX11-NEXT:    ; implicit-def: $sgpr46
; GFX11-NEXT:    ; implicit-def: $sgpr44
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    s_branch .LBB13_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) {
; SI-LABEL: bitcast_v128i8_to_v32i32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:188
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:204
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:212
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:220
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:268
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:292
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:316
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:332
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:352
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:376
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:348
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:384
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380
; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:364
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:372
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB14_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    v_or_b32_e32 v2, v2, v41
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_or_b32_e32 v3, v40, v3
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v2, v2, v53
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v48
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v62
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v38
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v56
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v46
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
; SI-NEXT:    v_or_b32_e32 v8, v8, v44
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v52
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v43
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:  .LBB14_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB14_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_movk_i32 s6, 0x300
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b32 s7, 0x3000000
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v54, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v40, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v41, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v53, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v48, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v62, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v38, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v56, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v46, v7
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v8, v44, v8
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v33
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v59
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v18
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v27
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
; SI-NEXT:    v_or_b32_e32 v31, v43, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v31
; SI-NEXT:  .LBB14_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v32i32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:12
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:380
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB14_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s6, 0xc0c0004
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v9, v33, v62, s6
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_perm_b32 v10, v63, v60, s6
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v11, v53, v58, s6
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_perm_b32 v12, v51, v56, s6
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v37, v36, s6
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v45, v44, s6
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v49, v48, s6
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v0, v1, v0, s6
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s6
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v41, s6
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_perm_b32 v8, v55, v34, s6
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_perm_b32 v9, v61, v32, s6
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_perm_b32 v10, v59, v54, s6
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_perm_b32 v11, v57, v52, s6
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v47, v38, s6
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_perm_b32 v13, v35, v46, s6
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_perm_b32 v14, v43, v50, s6
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_perm_b32 v15, v39, v42, s6
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v30, v30, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v32, v40, v32, s6
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:  .LBB14_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB14_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v31, 0x300
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v41
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v62
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v60
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v58
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v56
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v36
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v48
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    s_waitcnt vmcnt(11)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    s_waitcnt vmcnt(10)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    s_waitcnt vmcnt(9)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    s_waitcnt vmcnt(8)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    s_waitcnt vmcnt(7)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    s_waitcnt vmcnt(6)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_add_u16_e32 v8, 3, v55
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v34
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 3, v33
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_add_u16_e32 v9, 3, v61
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v32
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 3, v63
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
; VI-NEXT:    v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_add_u16_e32 v10, 3, v59
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v54
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 3, v53
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_add_u16_e32 v11, 3, v57
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v52
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 3, v51
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_add_u16_e32 v12, 3, v47
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 3, v37
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
; VI-NEXT:    v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_add_u16_e32 v13, 3, v35
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v46
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 3, v45
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_add_u16_e32 v14, 3, v43
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v50
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v15, 3, v49
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_add_u16_e32 v15, 3, v39
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v42
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v30, v32
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v33, 3, v40
; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:  .LBB14_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v32i32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:12
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:380
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB14_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0xc0c0004
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v10, v55, v32, s6
; GFX9-NEXT:    v_perm_b32 v11, v61, v60, s6
; GFX9-NEXT:    v_perm_b32 v12, v53, v38, s6
; GFX9-NEXT:    v_perm_b32 v13, v57, v56, s6
; GFX9-NEXT:    v_perm_b32 v14, v51, v34, s6
; GFX9-NEXT:    v_perm_b32 v15, v45, v50, s6
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v16, v43, v16, s6
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v9, s6
; GFX9-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX9-NEXT:    v_perm_b32 v8, v35, v42, s6
; GFX9-NEXT:    v_perm_b32 v9, v41, v48, s6
; GFX9-NEXT:    v_lshl_or_b32 v8, v9, 16, v8
; GFX9-NEXT:    v_perm_b32 v9, v39, v40, s6
; GFX9-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v10, v62, v63, s6
; GFX9-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX9-NEXT:    v_perm_b32 v11, v59, v54, s6
; GFX9-NEXT:    v_lshl_or_b32 v11, v12, 16, v11
; GFX9-NEXT:    v_perm_b32 v12, v58, v37, s6
; GFX9-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v47, v52, s6
; GFX9-NEXT:    v_lshl_or_b32 v13, v14, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v33, v46, s6
; GFX9-NEXT:    v_lshl_or_b32 v14, v15, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v49, v44, s6
; GFX9-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s6
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    v_lshl_or_b32 v16, v17, 16, v16
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    v_lshl_or_b32 v18, v19, 16, v18
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    v_lshl_or_b32 v19, v20, 16, v19
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    v_lshl_or_b32 v21, v22, 16, v21
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    v_lshl_or_b32 v23, v24, 16, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    v_lshl_or_b32 v24, v25, 16, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    v_lshl_or_b32 v25, v26, 16, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    v_lshl_or_b32 v26, v27, 16, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    v_lshl_or_b32 v28, v29, 16, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    v_lshl_or_b32 v29, v30, 16, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    v_lshl_or_b32 v30, v31, 16, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v32, v36, v32, s6
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    v_lshl_or_b32 v31, v32, 16, v31
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:  .LBB14_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB14_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    s_movk_i32 s6, 0x300
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v48
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v32
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v60
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v56
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v34
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v50
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(25)
; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(24)
; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
; GFX9-NEXT:    s_waitcnt vmcnt(22)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(19)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(14)
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(13)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    s_waitcnt vmcnt(11)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    s_waitcnt vmcnt(8)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v35
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v42
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v41
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v39
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v40
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v55
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v63
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v61
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v54
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v53
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v58
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v37
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v57
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v52
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v51
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v33
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v46
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v45
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v49
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v44
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v43
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v32, 3, v36
; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
; GFX9-NEXT:  .LBB14_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v32i32:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:384
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:380
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:376
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:372
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:368
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:364
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:360
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:356
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:352
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:348
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:344
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:340
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:336
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:332
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:328
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:260
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:132
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:388
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:8
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v74.l, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v75.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v76.l, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v77.l, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v78.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v88.l, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v79.l, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v89.l, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v90.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v92.l, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v91.l, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v93.l, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v94.l, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v95.l, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v104.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v106.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v105.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v107.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v108.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v110.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v109.l, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v111.l, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v120.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v121.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v122.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v123.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v124.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v125.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v126.l, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v127.l, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB14_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v127, v126, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v125, v124, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v123, v122, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v121, v120, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v107, v105, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v106, v104, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v95, v94, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v111, v109, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v110, v108, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v93, v91, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v92, v90, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v89, v79, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v88, v78, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v77, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v75, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v74, v72, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v63, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v62, v60, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v59, v58, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v57, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v56, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v45, v43, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v44, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v41, v40, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v183, v181, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v182, v180, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v179, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v178, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v164, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v161, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v160, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v149, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v147, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v146, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v135, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v134, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v131, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v129, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v128, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v117, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v116, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v113, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v103, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v102, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v99, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v98, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v87, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v84, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v81, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v80, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v69, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v66, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v55, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v54, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v48, v49, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v37, v39, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v38, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v36, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v48, 16, v31
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v35, 16, v37
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v32, 16, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr127_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr126_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr125_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr124_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr122_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr121_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr120_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr111_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr109_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr108_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr107_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr105_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr106_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr104_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr95_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr94_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr93_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr91_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr92_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr90_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr89_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr88_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT:  .LBB14_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB14_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v127.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v125.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v123.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v0.h, 8, v126.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v124.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v122.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v121.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v0.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v120.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v111.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v110.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v109.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v3.h, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v108.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, v107.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, v106.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v95.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v3.l, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v3.h, v2.h
; GFX11-TRUE16-NEXT:    v_and_b16 v3.l, 0xff, v4.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v105.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v4.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v104.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, v92.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v90.l
; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v3.h, v3.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, v89.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v94.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v93.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, v88.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v77.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v5.l, 8, v91.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, v74.l, 3
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v72.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, v63.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, v62.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v5.l, v4.h
; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v6.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v79.l
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v78.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v59.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, v56.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v76.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v75.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v10.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v46.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, v45.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.h, 8, v73.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, v44.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v41.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, v182.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v180.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v7.h, v7.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v9.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v61.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v9.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v60.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, v179.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, v178.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v58.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v57.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v167.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, v164.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v162.l
; GFX11-TRUE16-NEXT:    v_or_b16 v9.l, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v10.l, 8, v47.l
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, v161.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, v160.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v149.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v9.h, v10.l, v9.h
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v11.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v43.l
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v42.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, v146.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v144.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, v135.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.l, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v40.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v183.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, v134.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v131.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.h, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v12.l, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.h, 8, v181.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, v128.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v118.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, v117.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, v116.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.h, v12.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v14.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v177.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v14.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v176.l
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v113.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.l, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, v102.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.h, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v166.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v165.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v23.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v100.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, v99.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v15.l, 8, v163.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, v98.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v87.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, v84.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v82.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v15.l, v14.h
; GFX11-TRUE16-NEXT:    v_or_b16 v15.l, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v16.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v151.l
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v150.l
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, v81.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v15.h, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, v80.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v148.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v147.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v69.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, v66.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v28.h, 8, v64.l
; GFX11-TRUE16-NEXT:    v_or_b16 v16.h, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v17.l, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.h, 8, v145.l
; GFX11-TRUE16-NEXT:    v_and_b16 v28.l, 0xff, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, v55.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v29.h, 8, v53.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.l, 8, v51.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.l, v17.h, v17.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.h, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v19.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v133.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v19.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v132.l
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.h, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, v52.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v18.l, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v18.h, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v130.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v129.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v28.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v50.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.l, v29.h, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.l, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.h, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v20.l, 8, v119.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, v38.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, v36.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v33.h, v34.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.h, 8, v37.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.h, v20.l, v19.h
; GFX11-TRUE16-NEXT:    v_or_b16 v20.l, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v21.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v115.l
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v114.l
; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v31.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.l, 8, v35.l
; GFX11-TRUE16-NEXT:    v_or_b16 v20.h, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v21.l, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v112.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v103.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT:    v_and_b16 v33.h, 0xff, v33.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.l, 8, v32.l
; GFX11-TRUE16-NEXT:    v_or_b16 v21.h, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v22.l, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.h, 8, v101.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.l, v34.l, v31.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.h, v33.l, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v32.l, v33.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.l, v22.h, v22.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.h, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v24.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v97.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v24.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v96.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v23.l, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v23.h, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v86.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v85.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v25.l, 8, v83.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v3.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v4.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v4.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v5.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v25.l, v24.h
; GFX11-TRUE16-NEXT:    v_or_b16 v25.l, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v26.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v71.l
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v70.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v6.l
; GFX11-TRUE16-NEXT:    v_or_b16 v25.h, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v6.h
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v68.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v67.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v7.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v7.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v8.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.h, 8, v65.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v8.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v9.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v27.l, v27.h, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v54.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v11.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v11.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v12.l
; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v12.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v13.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v14.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.h, v30.l, v27.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v28.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, v49.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v29.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, v39.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v30.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v48.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v14.h
; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v15.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v15.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v16.l
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v32.h, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v16.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v17.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v17.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v18.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v19.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v19.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v20.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v21.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v21.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v22.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v22.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v23.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v24.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v24.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v25.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v26.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v26.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v30.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v31.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v31.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v32.l
; GFX11-TRUE16-NEXT:  .LBB14_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v32i32:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v86, v0
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:384
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:380
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:376
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:372
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:368
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:364
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:360
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:356
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:352
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:348
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:344
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:340
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:336
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:332
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:328
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:260
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:188
; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v45, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v46, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v47, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v57, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v56, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v59, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v61, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v62, off, s32 offset:132
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v63, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v73, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v72, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v74, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v75, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v79, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v89, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v88, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v90, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v91, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v93, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v94, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v95, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v105, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_u16 v104, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v106, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v107, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v109, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v108, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v110, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v111, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v120, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_u16 v121, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v122, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v0, off, s32 offset:388
; GFX11-FAKE16-NEXT:    scratch_load_u16 v123, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v125, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v124, off, s32 offset:8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_u16 v126, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v127, off, s32
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, v30 :: v_dual_mov_b32 v32, v29
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v28 :: v_dual_mov_b32 v35, v27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, v26 :: v_dual_mov_b32 v36, v25
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, v24 :: v_dual_mov_b32 v39, v23
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v49, v22 :: v_dual_mov_b32 v48, v21
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v16 :: v_dual_mov_b32 v55, v15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v65, v14 :: v_dual_mov_b32 v64, v13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, v12 :: v_dual_mov_b32 v67, v11
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, v10 :: v_dual_mov_b32 v68, v9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, v8 :: v_dual_mov_b32 v71, v7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, v4 :: v_dual_mov_b32 v83, v3
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, v2 :: v_dual_mov_b32 v84, v1
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB14_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v86, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v85, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v80, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v66, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v65, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v54, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v69, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v52, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v50, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v49, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v38, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v37, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v33, v127, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v126, v124, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v125, v123, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v122, v121, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v120, v111, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v110, v108, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v109, v107, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v106, v104, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v105, v95, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v94, v92, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v93, v91, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v90, v88, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v89, v79, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v78, v77, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v76, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v74, v72, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v73, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v62, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v61, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v58, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v57, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v46, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v45, v43, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v42, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v40, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v182, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v181, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v178, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v177, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v166, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v165, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v162, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v161, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v150, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v148, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v146, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v145, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v134, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v133, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v130, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v129, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v118, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v117, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v115, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v112, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v103, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v34, v101, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v99, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v97, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v31
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v34, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v36, 16, v35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr127
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr126
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr124
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr125
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr122
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr121
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr120
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr111
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr108
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr109
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr107
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr106
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr105
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr95
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr93
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr91
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr89
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:  .LBB14_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB14_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v86, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v85, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v84
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v82, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v80, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v81
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v69, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v70, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v67
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v68
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v66, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v54, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v65, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v55
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v7, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v52, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v51
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v50, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v49, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v8, 8, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, v38, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, v37, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v34, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v33, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v10, 8, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v127
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v126, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v125, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v11, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v124
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v123
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v122, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v121
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v120, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v111
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v110, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v109, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v107
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v13, 8, v108
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, v106, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, v105, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v14, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v104
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v95
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v94, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v12, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v93, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v15, 8, v92
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v91
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v90, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v89, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v88
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v79
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v78, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v17, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v77
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v76, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v14, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v75
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v74, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v73, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v63
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v18, 8, v72
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, v62, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, v61, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v19, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v60
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v59
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v58, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v57, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v20, 8, v56
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v47
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v46, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v45, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v21, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v44
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v43
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v42, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v22, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v41
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v40, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v183
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v182, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v181, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v179
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v23, 8, v180
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, v178, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, v177, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v24, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v176
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v167
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v166, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v22, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v165, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v25, 8, v164
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v163
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v162, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v161, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v26, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v160
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v151
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v150, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v27, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v149
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v148, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v146, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v145, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v135
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v28, 8, v144
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v134, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v133, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v29, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v132
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v131
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, v130, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v129, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v30, 8, v128
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v119
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v26, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v118, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v117, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v31, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v115, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v116
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v33, 8, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v113
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v32, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, v112, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, v103, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, v101, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, v99, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v38, v97, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v34, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v102
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v37, 8, v100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v39, 8, v98
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v48, 8, v96
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v49, 8, v87
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v34, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v37, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v39, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v48, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v49, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v35, v36
; GFX11-FAKE16-NEXT:  .LBB14_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define inreg <32 x i32> @bitcast_v128i8_to_v32i32_scalar(<128 x i8> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v128i8_to_v32i32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s76, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s46, v30
; SI-NEXT:    v_readfirstlane_b32 s6, v29
; SI-NEXT:    v_readfirstlane_b32 s7, v28
; SI-NEXT:    v_readfirstlane_b32 s8, v27
; SI-NEXT:    v_readfirstlane_b32 s9, v26
; SI-NEXT:    v_readfirstlane_b32 s10, v25
; SI-NEXT:    v_readfirstlane_b32 s11, v24
; SI-NEXT:    v_readfirstlane_b32 s12, v23
; SI-NEXT:    v_readfirstlane_b32 s13, v22
; SI-NEXT:    v_readfirstlane_b32 s14, v21
; SI-NEXT:    v_readfirstlane_b32 s15, v20
; SI-NEXT:    v_readfirstlane_b32 s40, v19
; SI-NEXT:    v_readfirstlane_b32 s41, v18
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:96
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:104
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:328
; SI-NEXT:    v_readfirstlane_b32 s75, v1
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:332
; SI-NEXT:    v_readfirstlane_b32 s43, v16
; SI-NEXT:    v_readfirstlane_b32 s45, v14
; SI-NEXT:    v_readfirstlane_b32 s73, v3
; SI-NEXT:    v_readfirstlane_b32 s74, v2
; SI-NEXT:    v_readfirstlane_b32 s57, v11
; SI-NEXT:    v_readfirstlane_b32 s72, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v15
; SI-NEXT:    v_readfirstlane_b32 s47, v13
; SI-NEXT:    v_readfirstlane_b32 s56, v12
; SI-NEXT:    v_readfirstlane_b32 s58, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s60, v8
; SI-NEXT:    v_readfirstlane_b32 s61, v7
; SI-NEXT:    v_readfirstlane_b32 s62, v6
; SI-NEXT:    v_readfirstlane_b32 s63, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v50
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v51
; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v52
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v13, 24, v54
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v55
; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v40
; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v41
; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v42
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v43
; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v46
; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v56
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v59
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v61
; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v44
; SI-NEXT:    v_lshlrev_b32_e32 v52, 8, v45
; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v47
; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:316
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v30
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 24, v29
; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v28
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v26
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v25
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v24
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v23
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v22
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v20
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v19
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v18
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:268
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:260
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:252
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:292
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v34
; SI-NEXT:    v_lshlrev_b32_e32 v34, 8, v32
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:220
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204
; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 24, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v33
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:196
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:116
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB15_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v31
; SI-NEXT:    v_or_b32_e32 v0, s4, v0
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v12
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v11, v3
; SI-NEXT:    v_or_b32_e32 v11, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v30
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v13, v1
; SI-NEXT:    v_or_b32_e32 v13, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v9
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v14, v1
; SI-NEXT:    v_or_b32_e32 v14, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v42
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v0, v0, v60
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v15, v1
; SI-NEXT:    v_or_b32_e32 v15, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v63
; SI-NEXT:    v_or_b32_e32 v0, v0, v41
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v16, v1
; SI-NEXT:    v_or_b32_e32 v16, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v61
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v17
; SI-NEXT:    v_or_b32_e32 v0, v0, v40
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v55, v1
; SI-NEXT:    v_or_b32_e32 v17, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v44
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v53, v1
; SI-NEXT:    v_or_b32_e32 v18, v0, v1
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v51, v1
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v12, v2, v3
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s18, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s77, s19, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s77, s5
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s77, s21, 8
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s22, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s77, s78, s77
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s24, 0xff
; SI-NEXT:    s_lshl_b32 s78, s25, 8
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s26, 0xff
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s79, s27, 24
; SI-NEXT:    s_and_b32 s77, s77, 0xffff
; SI-NEXT:    s_or_b32 s78, s79, s78
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s28, 0xff
; SI-NEXT:    s_lshl_b32 s79, s29, 8
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s76, 0xff
; SI-NEXT:    s_lshl_b32 s79, s79, 16
; SI-NEXT:    s_lshl_b32 s88, s75, 24
; SI-NEXT:    s_and_b32 s78, s78, 0xffff
; SI-NEXT:    s_or_b32 s79, s88, s79
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s74, 0xff
; SI-NEXT:    s_lshl_b32 s88, s73, 8
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s72, 0xff
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s89, s63, 24
; SI-NEXT:    s_and_b32 s79, s79, 0xffff
; SI-NEXT:    s_or_b32 s88, s89, s88
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s62, 0xff
; SI-NEXT:    s_lshl_b32 s89, s61, 8
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s60, 0xff
; SI-NEXT:    s_lshl_b32 s89, s89, 16
; SI-NEXT:    s_lshl_b32 s90, s59, 24
; SI-NEXT:    s_and_b32 s88, s88, 0xffff
; SI-NEXT:    s_or_b32 s89, s90, s89
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s58, 0xff
; SI-NEXT:    s_lshl_b32 s90, s57, 8
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s56, 0xff
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_lshl_b32 s91, s47, 24
; SI-NEXT:    s_and_b32 s89, s89, 0xffff
; SI-NEXT:    s_or_b32 s90, s91, s90
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s45, 0xff
; SI-NEXT:    s_lshl_b32 s91, s44, 8
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s43, 0xff
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v52
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v19, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v49, v1
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    s_lshl_b32 s91, s91, 16
; SI-NEXT:    s_lshl_b32 s92, s42, 24
; SI-NEXT:    s_and_b32 s90, s90, 0xffff
; SI-NEXT:    s_or_b32 s91, s92, s91
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s41, 0xff
; SI-NEXT:    s_lshl_b32 s92, s40, 8
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s15, 0xff
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s93, s14, 24
; SI-NEXT:    s_and_b32 s91, s91, 0xffff
; SI-NEXT:    s_or_b32 s92, s93, s92
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s13, 0xff
; SI-NEXT:    s_lshl_b32 s93, s12, 8
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s11, 0xff
; SI-NEXT:    s_lshl_b32 s93, s93, 16
; SI-NEXT:    s_lshl_b32 s94, s10, 24
; SI-NEXT:    s_and_b32 s92, s92, 0xffff
; SI-NEXT:    s_or_b32 s93, s94, s93
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s9, 0xff
; SI-NEXT:    s_lshl_b32 s94, s8, 8
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    s_and_b32 s94, s7, 0xff
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_lshl_b32 s95, s6, 24
; SI-NEXT:    s_and_b32 s93, s93, 0xffff
; SI-NEXT:    s_or_b32 s94, s95, s94
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v3, s78
; SI-NEXT:    v_mov_b32_e32 v4, s79
; SI-NEXT:    v_mov_b32_e32 v5, s88
; SI-NEXT:    v_mov_b32_e32 v6, s89
; SI-NEXT:    v_mov_b32_e32 v7, s90
; SI-NEXT:    v_mov_b32_e32 v8, s91
; SI-NEXT:    v_mov_b32_e32 v9, s92
; SI-NEXT:    v_mov_b32_e32 v10, s93
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v52
; SI-NEXT:    v_or_b32_e32 v0, v0, v50
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v20, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v39, v1
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v48
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v21, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v37, v1
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v48
; SI-NEXT:    v_or_b32_e32 v0, v0, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v22, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v35, v1
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v38
; SI-NEXT:    v_or_b32_e32 v0, v0, v36
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v23, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v33, v1
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v36
; SI-NEXT:    v_or_b32_e32 v0, v0, v34
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v24, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    v_or_b32_e32 v0, v0, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v29, v1
; SI-NEXT:    v_or_b32_e32 v25, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v28, v1
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v0, v0, v59
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v26, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v58
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v56
; SI-NEXT:    v_or_b32_e32 v0, v0, v57
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v27, v1
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    v_or_b32_e32 v27, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v0, v0, v45
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_or_b32_e32 v28, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v46
; SI-NEXT:    v_mov_b32_e32 v45, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v60, v32
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v29, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v30, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v31, v0, v1
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s77
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_branch .LBB15_3
; SI-NEXT:  .LBB15_2:
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB15_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v43, v58
; SI-NEXT:    v_mov_b32_e32 v58, v33
; SI-NEXT:    v_mov_b32_e32 v33, v37
; SI-NEXT:    v_mov_b32_e32 v34, v39
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_cbranch_vccnz .LBB15_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_mov_b32_e32 v32, v35
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_and_b32 s17, s22, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s76, s76, 3
; SI-NEXT:    s_add_i32 s74, s74, 3
; SI-NEXT:    s_add_i32 s72, s72, 3
; SI-NEXT:    s_add_i32 s62, s62, 3
; SI-NEXT:    s_add_i32 s60, s60, 3
; SI-NEXT:    s_add_i32 s58, s58, 3
; SI-NEXT:    s_add_i32 s56, s56, 3
; SI-NEXT:    s_and_b32 s22, s56, 0xff
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_addk_i32 s12, 0x300
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    s_add_i32 s10, s10, 0x3000000
; SI-NEXT:    s_add_i32 s6, s6, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v9, s10
; SI-NEXT:    v_mov_b32_e32 v10, s6
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, s4, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, 0x3000000, v0
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s16, s18, 0xff
; SI-NEXT:    s_addk_i32 s4, 0x300
; SI-NEXT:    s_lshl_b32 s5, s19, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s5, s16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s16, s21, 8
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_addk_i32 s5, 0x300
; SI-NEXT:    s_lshl_b32 s16, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_and_b32 s16, s24, 0xff
; SI-NEXT:    s_lshl_b32 s17, s25, 8
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s18, s26, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s27, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s28, 0xff
; SI-NEXT:    s_lshl_b32 s18, s29, 8
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s19, s76, 0xff
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_lshl_b32 s18, s75, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s18, s74, 0xff
; SI-NEXT:    s_lshl_b32 s19, s73, 8
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s20, s72, 0xff
; SI-NEXT:    s_addk_i32 s18, 0x300
; SI-NEXT:    s_lshl_b32 s19, s63, 24
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s19, s19, s20
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s19, s62, 0xff
; SI-NEXT:    s_lshl_b32 s20, s61, 8
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s21, s60, 0xff
; SI-NEXT:    s_addk_i32 s19, 0x300
; SI-NEXT:    s_lshl_b32 s20, s59, 24
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s21
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s20, s58, 0xff
; SI-NEXT:    s_lshl_b32 s21, s57, 8
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_addk_i32 s20, 0x300
; SI-NEXT:    s_lshl_b32 s21, s47, 24
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_and_b32 s21, s45, 0xff
; SI-NEXT:    s_lshl_b32 s22, s44, 8
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    s_and_b32 s23, s43, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s22, s42, 24
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s23
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s22, s41, 0xff
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_lshl_b32 s23, s40, 8
; SI-NEXT:    s_or_b32 s22, s23, s22
; SI-NEXT:    s_addk_i32 s22, 0x300
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s14, s14, s22
; SI-NEXT:    s_add_i32 s4, s4, 0x3000000
; SI-NEXT:    s_add_i32 s5, s5, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 0x3000000
; SI-NEXT:    s_add_i32 s17, s17, 0x3000000
; SI-NEXT:    s_add_i32 s18, s18, 0x3000000
; SI-NEXT:    s_add_i32 s19, s19, 0x3000000
; SI-NEXT:    s_add_i32 s20, s20, 0x3000000
; SI-NEXT:    s_add_i32 s21, s21, 0x3000000
; SI-NEXT:    s_add_i32 s14, s14, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v4, s18
; SI-NEXT:    v_mov_b32_e32 v5, s19
; SI-NEXT:    v_mov_b32_e32 v6, s20
; SI-NEXT:    v_mov_b32_e32 v7, s21
; SI-NEXT:    v_mov_b32_e32 v8, s14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_add_i32_e32 v12, vcc, 0x3000000, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v39, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_mov_b32_e32 v3, s17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v13, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v37, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v42, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v41, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v44
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v55, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v54
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v53, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v51, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v48
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v34, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v38
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v33, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v32, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v58
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v62, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v60
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v59, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v26, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v57, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v27, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v45, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v28, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v0
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:  .LBB15_5: ; %end
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v32i32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s45, v15
; VI-NEXT:    v_readfirstlane_b32 s46, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s76, v0
; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_readfirstlane_b32 s6, v30
; VI-NEXT:    v_readfirstlane_b32 s7, v29
; VI-NEXT:    v_readfirstlane_b32 s8, v28
; VI-NEXT:    v_readfirstlane_b32 s9, v27
; VI-NEXT:    v_readfirstlane_b32 s10, v26
; VI-NEXT:    v_readfirstlane_b32 s11, v25
; VI-NEXT:    v_readfirstlane_b32 s12, v24
; VI-NEXT:    v_readfirstlane_b32 s13, v23
; VI-NEXT:    v_readfirstlane_b32 s14, v22
; VI-NEXT:    v_readfirstlane_b32 s15, v21
; VI-NEXT:    v_readfirstlane_b32 s40, v20
; VI-NEXT:    v_readfirstlane_b32 s41, v19
; VI-NEXT:    v_readfirstlane_b32 s42, v18
; VI-NEXT:    v_readfirstlane_b32 s43, v17
; VI-NEXT:    v_readfirstlane_b32 s44, v16
; VI-NEXT:    v_readfirstlane_b32 s56, v12
; VI-NEXT:    v_readfirstlane_b32 s57, v11
; VI-NEXT:    v_readfirstlane_b32 s58, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s60, v8
; VI-NEXT:    v_readfirstlane_b32 s61, v7
; VI-NEXT:    v_readfirstlane_b32 s62, v6
; VI-NEXT:    v_readfirstlane_b32 s63, v5
; VI-NEXT:    v_readfirstlane_b32 s72, v4
; VI-NEXT:    v_readfirstlane_b32 s73, v3
; VI-NEXT:    v_readfirstlane_b32 s74, v2
; VI-NEXT:    v_readfirstlane_b32 s75, v1
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; VI-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; VI-NEXT:    s_and_b64 s[4:5], vcc, exec
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    s_cbranch_scc0 .LBB15_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_mov_b32_e32 v3, s23
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_mov_b32_e32 v2, s21
; VI-NEXT:    v_perm_b32 v3, s22, v3, v11
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_perm_b32 v2, s20, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_or_b32_e32 v10, v10, v12
; VI-NEXT:    v_perm_b32 v12, v14, v13, s4
; VI-NEXT:    v_perm_b32 v11, s6, v15, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v52, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v54, v53, s4
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    v_perm_b32 v13, v48, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v50, v49, s4
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_perm_b32 v14, v36, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v38, v37, s4
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_perm_b32 v15, v32, v16, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v16, v34, v33, s4
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    v_perm_b32 v16, v60, v17, s4
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_perm_b32 v17, v62, v61, s4
; VI-NEXT:    v_or_b32_e32 v16, v17, v16
; VI-NEXT:    v_perm_b32 v17, v59, v18, s4
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_perm_b32 v18, v41, v40, s4
; VI-NEXT:    v_or_b32_e32 v17, v18, v17
; VI-NEXT:    v_perm_b32 v18, v57, v19, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v19, v42, v58, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_perm_b32 v19, v45, v20, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_perm_b32 v20, v56, v47, s4
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_perm_b32 v20, v44, v21, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_perm_b32 v21, v46, v26, s4
; VI-NEXT:    v_or_b32_e32 v20, v21, v20
; VI-NEXT:    v_perm_b32 v21, v31, v22, s4
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_perm_b32 v22, v24, v43, s4
; VI-NEXT:    v_or_b32_e32 v21, v22, v21
; VI-NEXT:    v_perm_b32 v22, v28, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_perm_b32 v23, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v22, v23, v22
; VI-NEXT:    v_perm_b32 v23, v25, v63, s4
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_perm_b32 v24, v27, v55, s4
; VI-NEXT:    v_or_b32_e32 v23, v24, v23
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    v_or_b32_e32 v24, v25, v24
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    v_or_b32_e32 v25, v26, v25
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    v_or_b32_e32 v26, v27, v26
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    v_or_b32_e32 v27, v28, v27
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v28, v29, v28
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    v_or_b32_e32 v29, v30, v29
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    v_or_b32_e32 v30, v31, v30
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v43, v46, v43, s4
; VI-NEXT:    v_or_b32_e32 v31, v43, v31
; VI-NEXT:    s_mov_b64 s[4:5], 0
; VI-NEXT:    s_branch .LBB15_3
; VI-NEXT:  .LBB15_2:
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    s_mov_b64 s[4:5], -1
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:  .LBB15_3: ; %Flow
; VI-NEXT:    v_mov_b32_e32 v43, v44
; VI-NEXT:    v_mov_b32_e32 v44, v45
; VI-NEXT:    v_mov_b32_e32 v45, v47
; VI-NEXT:    v_mov_b32_e32 v46, v57
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; VI-NEXT:    s_cbranch_vccnz .LBB15_5
; VI-NEXT:  ; %bb.4: ; %cmp.true
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    v_mov_b32_e32 v1, s21
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    v_mov_b32_e32 v2, s23
; VI-NEXT:    v_perm_b32 v1, s20, v1, v11
; VI-NEXT:    v_perm_b32 v2, s22, v2, v11
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x300, v2
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    s_add_i32 s76, s76, 3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x300, v3
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s74, s74, 3
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    s_add_i32 s72, s72, 3
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x300, v4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s62, s62, 3
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    s_add_i32 s60, s60, 3
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x300, v5
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s58, s58, 3
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    s_add_i32 s56, s56, 3
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x300, v6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x300, v7
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x300, v8
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x300, v9
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x300, v10
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v10, v12, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v48
; VI-NEXT:    v_perm_b32 v14, v14, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v36
; VI-NEXT:    v_perm_b32 v15, v15, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x3000000, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x3000000, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x3000000, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v6
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v7
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x3000000, v8
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x3000000, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v10
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v11, s6, v12, v11
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x300, v11
; VI-NEXT:    v_perm_b32 v16, v16, v17, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v60
; VI-NEXT:    v_perm_b32 v17, v17, v18, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v57
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v18, v18, v19, s4
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v12, v12, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v54
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v52
; VI-NEXT:    v_perm_b32 v12, v12, v53, s4
; VI-NEXT:    v_perm_b32 v13, v13, v51, s4
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x300, v12
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v50
; VI-NEXT:    v_perm_b32 v13, v13, v49, s4
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x300, v13
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v38
; VI-NEXT:    v_perm_b32 v14, v14, v37, s4
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x300, v14
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v34
; VI-NEXT:    v_perm_b32 v15, v15, v33, s4
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x300, v15
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v62
; VI-NEXT:    v_perm_b32 v16, v16, v61, s4
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x300, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v41
; VI-NEXT:    v_perm_b32 v17, v17, v40, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x300, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v42
; VI-NEXT:    v_perm_b32 v18, v18, v47, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v18
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v46
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x3000000, v11
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x3000000, v12
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x3000000, v13
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x3000000, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x3000000, v15
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x3000000, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x3000000, v17
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x3000000, v18
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_perm_b32 v20, v20, v21, s4
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v45
; VI-NEXT:    v_perm_b32 v20, v20, v44, s4
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x300, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x3000000, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v43
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x3000000, v21
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x3000000, v22
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x3000000, v23
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x3000000, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x3000000, v25
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x3000000, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x3000000, v27
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x3000000, v28
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x3000000, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x3000000, v30
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x300, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v32
; VI-NEXT:    v_perm_b32 v32, v32, v33, s4
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x3000000, v31
; VI-NEXT:  .LBB15_5: ; %end
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v32i32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s45, v15
; GFX9-NEXT:    v_readfirstlane_b32 s46, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s76, v0
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readfirstlane_b32 s6, v30
; GFX9-NEXT:    v_readfirstlane_b32 s7, v29
; GFX9-NEXT:    v_readfirstlane_b32 s8, v28
; GFX9-NEXT:    v_readfirstlane_b32 s9, v27
; GFX9-NEXT:    v_readfirstlane_b32 s10, v26
; GFX9-NEXT:    v_readfirstlane_b32 s11, v25
; GFX9-NEXT:    v_readfirstlane_b32 s12, v24
; GFX9-NEXT:    v_readfirstlane_b32 s13, v23
; GFX9-NEXT:    v_readfirstlane_b32 s14, v22
; GFX9-NEXT:    v_readfirstlane_b32 s15, v21
; GFX9-NEXT:    v_readfirstlane_b32 s40, v20
; GFX9-NEXT:    v_readfirstlane_b32 s41, v19
; GFX9-NEXT:    v_readfirstlane_b32 s42, v18
; GFX9-NEXT:    v_readfirstlane_b32 s43, v17
; GFX9-NEXT:    v_readfirstlane_b32 s44, v16
; GFX9-NEXT:    v_readfirstlane_b32 s56, v12
; GFX9-NEXT:    v_readfirstlane_b32 s57, v11
; GFX9-NEXT:    v_readfirstlane_b32 s58, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s60, v8
; GFX9-NEXT:    v_readfirstlane_b32 s61, v7
; GFX9-NEXT:    v_readfirstlane_b32 s62, v6
; GFX9-NEXT:    v_readfirstlane_b32 s63, v5
; GFX9-NEXT:    v_readfirstlane_b32 s72, v4
; GFX9-NEXT:    v_readfirstlane_b32 s73, v3
; GFX9-NEXT:    v_readfirstlane_b32 s74, v2
; GFX9-NEXT:    v_readfirstlane_b32 s75, v1
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    s_cbranch_scc0 .LBB15_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v12
; GFX9-NEXT:    v_perm_b32 v12, v14, v13, s4
; GFX9-NEXT:    v_perm_b32 v11, s6, v15, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_perm_b32 v12, v52, v51, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v54, v53, s4
; GFX9-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX9-NEXT:    v_perm_b32 v13, v48, v39, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v50, v49, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX9-NEXT:    v_perm_b32 v14, v36, v35, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v38, v37, s4
; GFX9-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX9-NEXT:    v_perm_b32 v15, v32, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    v_perm_b32 v16, v34, v33, s4
; GFX9-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX9-NEXT:    v_perm_b32 v16, v60, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_perm_b32 v17, v62, v61, s4
; GFX9-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX9-NEXT:    v_perm_b32 v17, v59, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_perm_b32 v18, v41, v40, s4
; GFX9-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX9-NEXT:    v_perm_b32 v18, v57, v19, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_perm_b32 v19, v42, v58, s4
; GFX9-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX9-NEXT:    v_perm_b32 v19, v45, v20, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX9-NEXT:    v_perm_b32 v20, v56, v47, s4
; GFX9-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX9-NEXT:    v_perm_b32 v20, v44, v21, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX9-NEXT:    v_perm_b32 v21, v46, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX9-NEXT:    v_perm_b32 v21, v31, v22, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX9-NEXT:    v_perm_b32 v22, v24, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX9-NEXT:    v_perm_b32 v22, v28, v23, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX9-NEXT:    v_perm_b32 v23, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX9-NEXT:    v_perm_b32 v23, v25, v63, s4
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX9-NEXT:    v_perm_b32 v24, v27, v55, s4
; GFX9-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s4
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    v_or_b32_e32 v30, v31, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v43, v46, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v31, v43, v31
; GFX9-NEXT:    s_mov_b64 s[4:5], 0
; GFX9-NEXT:    s_branch .LBB15_3
; GFX9-NEXT:  .LBB15_2:
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    s_mov_b64 s[4:5], -1
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:  .LBB15_3: ; %Flow
; GFX9-NEXT:    v_mov_b32_e32 v43, v44
; GFX9-NEXT:    v_mov_b32_e32 v44, v45
; GFX9-NEXT:    v_mov_b32_e32 v45, v47
; GFX9-NEXT:    v_mov_b32_e32 v46, v57
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_vccnz .LBB15_5
; GFX9-NEXT:  ; %bb.4: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    s_movk_i32 s4, 0x300
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    s_add_i32 s76, s76, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_add_u32_e32 v3, 0x300, v3
; GFX9-NEXT:    v_add_u32_sdwa v4, v4, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s74, s74, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    s_add_i32 s72, s72, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_add_u32_e32 v4, 0x300, v4
; GFX9-NEXT:    v_add_u32_sdwa v5, v5, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s62, s62, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    s_add_i32 s60, s60, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_add_u32_e32 v5, 0x300, v5
; GFX9-NEXT:    v_add_u32_sdwa v6, v6, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s58, s58, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    s_add_i32 s56, s56, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_add_u32_e32 v6, 0x300, v6
; GFX9-NEXT:    v_add_u32_sdwa v7, v7, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_add_u32_e32 v7, 0x300, v7
; GFX9-NEXT:    v_add_u32_sdwa v8, v8, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_add_u32_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u32_sdwa v9, v9, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_add_u32_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u32_sdwa v10, v10, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s5, 0xc0c0004
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v32
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v48
; GFX9-NEXT:    v_perm_b32 v14, v14, v39, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v14, v14, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v36
; GFX9-NEXT:    v_perm_b32 v15, v15, v35, s5
; GFX9-NEXT:    v_add_u32_sdwa v15, v15, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v60
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_perm_b32 v17, v17, v18, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v57
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v18, v18, v19, s5
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v11, s6, v12, v11
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u32_sdwa v16, v16, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v17, v17, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v18, v18, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_perm_b32 v12, v12, v13, s5
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v54
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v52
; GFX9-NEXT:    v_perm_b32 v12, v12, v53, s5
; GFX9-NEXT:    v_perm_b32 v13, v13, v51, s5
; GFX9-NEXT:    v_add_u32_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u32_sdwa v13, v13, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v50
; GFX9-NEXT:    v_perm_b32 v13, v13, v49, s5
; GFX9-NEXT:    v_add_u32_e32 v13, 0x300, v13
; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v38
; GFX9-NEXT:    v_perm_b32 v14, v14, v37, s5
; GFX9-NEXT:    v_add_u32_e32 v14, 0x300, v14
; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v34
; GFX9-NEXT:    v_perm_b32 v15, v15, v33, s5
; GFX9-NEXT:    v_add_u32_e32 v15, 0x300, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v62
; GFX9-NEXT:    v_perm_b32 v16, v16, v61, s5
; GFX9-NEXT:    v_add_u32_e32 v16, 0x300, v16
; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v41
; GFX9-NEXT:    v_perm_b32 v17, v17, v40, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 0x300, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v42
; GFX9-NEXT:    v_perm_b32 v18, v18, v47, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 0x300, v18
; GFX9-NEXT:    v_add_u32_sdwa v19, v19, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v46
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_perm_b32 v20, v20, v21, s5
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v20, v20, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v45
; GFX9-NEXT:    v_perm_b32 v20, v20, v44, s5
; GFX9-NEXT:    v_add_u32_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v21, v21, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    v_add_u32_sdwa v22, v22, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    v_add_u32_sdwa v23, v23, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    v_add_u32_sdwa v24, v24, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    v_add_u32_sdwa v25, v25, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    v_add_u32_sdwa v26, v26, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    v_add_u32_sdwa v27, v27, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    v_add_u32_sdwa v28, v28, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    v_add_u32_sdwa v29, v29, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    v_add_u32_sdwa v30, v30, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    v_add_u32_sdwa v31, v31, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v32, 3, v32
; GFX9-NEXT:    v_perm_b32 v32, v32, v33, s5
; GFX9-NEXT:    v_add_u32_sdwa v32, v32, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:  .LBB15_5: ; %end
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v32i32_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:260
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:188
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:60
; GFX11-TRUE16-NEXT:    s_clause 0xf
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s75, 0
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB15_4
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v135, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v55, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v147, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v99, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v103, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v67, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB15_3
; GFX11-TRUE16-NEXT:  .LBB15_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v135
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v99
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v55
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-TRUE16-NEXT:  .LBB15_3: ; %end
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT:  .LBB15_4:
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_branch .LBB15_2
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v32i32_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v32, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v34, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v33, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v37, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v35, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v39, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v36, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v48, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v38, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v49, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v50, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v52, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v51, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:260
; GFX11-FAKE16-NEXT:    scratch_load_u16 v53, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v54, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v55, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:188
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:132
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:60
; GFX11-FAKE16-NEXT:    s_clause 0xf
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s75, 0
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-FAKE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB15_4
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v144, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v64, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v147, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v103, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v67, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB15_3
; GFX11-FAKE16-NEXT:  .LBB15_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v144
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v100
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v64
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-FAKE16-NEXT:  .LBB15_3: ; %end
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-FAKE16-NEXT:  .LBB15_4:
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_branch .LBB15_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define <64 x bfloat> @bitcast_v32i32_to_v64bf16(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v64bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB16_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v62
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v62
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v13
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v11
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v10
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v9
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v8
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v7
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v3
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v2
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:  .LBB16_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB16_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v62
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v13
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v11
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v10
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v9
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v8
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v7
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v3
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v2
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:  .LBB16_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v60
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v58
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v56
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v46
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v44
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v42
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v40
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v54
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v52
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v50
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v48
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v39
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v38
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v37
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v36
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v35
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v34
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v33
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v32
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_alignbit_b32 v18, v18, v19, 16
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_alignbit_b32 v19, v19, v20, 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_alignbit_b32 v21, v21, v22, 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_alignbit_b32 v22, v22, v23, 16
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_alignbit_b32 v23, v23, v24, 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_alignbit_b32 v24, v24, v25, 16
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_alignbit_b32 v25, v25, v26, 16
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_alignbit_b32 v26, v26, v27, 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_alignbit_b32 v27, v27, v28, 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_alignbit_b32 v28, v28, v29, 16
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_alignbit_b32 v29, v29, v30, 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_alignbit_b32 v30, v30, v31, 16
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v63
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v64bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB16_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:  .LBB16_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32i32_to_v64bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB16_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11
; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10
; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9
; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8
; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7
; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6
; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5
; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4
; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3
; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16
; GFX9-NEXT:  .LBB16_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32i32_to_v64bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB16_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-NEXT:  .LBB16_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define inreg <64 x bfloat> @bitcast_v32i32_to_v64bf16_scalar(<32 x i32> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32i32_to_v64bf16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v33, s30, 0
; SI-NEXT:    v_writelane_b32 v33, s31, 1
; SI-NEXT:    v_writelane_b32 v33, s34, 2
; SI-NEXT:    v_writelane_b32 v33, s35, 3
; SI-NEXT:    v_writelane_b32 v33, s36, 4
; SI-NEXT:    v_writelane_b32 v33, s37, 5
; SI-NEXT:    v_writelane_b32 v33, s38, 6
; SI-NEXT:    v_writelane_b32 v33, s39, 7
; SI-NEXT:    v_writelane_b32 v33, s48, 8
; SI-NEXT:    v_writelane_b32 v33, s49, 9
; SI-NEXT:    v_writelane_b32 v33, s50, 10
; SI-NEXT:    v_writelane_b32 v33, s51, 11
; SI-NEXT:    v_writelane_b32 v33, s52, 12
; SI-NEXT:    v_writelane_b32 v33, s53, 13
; SI-NEXT:    v_writelane_b32 v33, s54, 14
; SI-NEXT:    v_writelane_b32 v33, s55, 15
; SI-NEXT:    v_writelane_b32 v33, s64, 16
; SI-NEXT:    v_writelane_b32 v33, s65, 17
; SI-NEXT:    v_writelane_b32 v33, s66, 18
; SI-NEXT:    v_writelane_b32 v33, s67, 19
; SI-NEXT:    v_writelane_b32 v33, s68, 20
; SI-NEXT:    v_writelane_b32 v33, s69, 21
; SI-NEXT:    v_writelane_b32 v33, s70, 22
; SI-NEXT:    v_writelane_b32 v33, s71, 23
; SI-NEXT:    v_writelane_b32 v33, s80, 24
; SI-NEXT:    v_writelane_b32 v33, s81, 25
; SI-NEXT:    v_writelane_b32 v33, s82, 26
; SI-NEXT:    v_writelane_b32 v33, s83, 27
; SI-NEXT:    v_writelane_b32 v33, s84, 28
; SI-NEXT:    v_writelane_b32 v33, s85, 29
; SI-NEXT:    v_writelane_b32 v33, s86, 30
; SI-NEXT:    v_writelane_b32 v33, s87, 31
; SI-NEXT:    v_writelane_b32 v33, s96, 32
; SI-NEXT:    v_writelane_b32 v33, s97, 33
; SI-NEXT:    v_writelane_b32 v33, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v33, s99, 35
; SI-NEXT:    v_readfirstlane_b32 s70, v17
; SI-NEXT:    v_readfirstlane_b32 s71, v16
; SI-NEXT:    v_readfirstlane_b32 s80, v15
; SI-NEXT:    v_readfirstlane_b32 s81, v14
; SI-NEXT:    v_readfirstlane_b32 s82, v13
; SI-NEXT:    v_readfirstlane_b32 s83, v12
; SI-NEXT:    v_readfirstlane_b32 s84, v11
; SI-NEXT:    v_readfirstlane_b32 s85, v10
; SI-NEXT:    v_readfirstlane_b32 s86, v9
; SI-NEXT:    v_readfirstlane_b32 s87, v8
; SI-NEXT:    v_readfirstlane_b32 s96, v7
; SI-NEXT:    v_readfirstlane_b32 s97, v6
; SI-NEXT:    v_readfirstlane_b32 s98, v5
; SI-NEXT:    v_readfirstlane_b32 s99, v4
; SI-NEXT:    v_readfirstlane_b32 s6, v3
; SI-NEXT:    v_readfirstlane_b32 s7, v2
; SI-NEXT:    v_readfirstlane_b32 s8, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s9, v0
; SI-NEXT:    ; implicit-def: $vgpr34 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB17_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s17, 0xffff0000
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v34, s4, 1
; SI-NEXT:    s_lshl_b32 s4, s17, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 0
; SI-NEXT:    s_and_b32 s4, s16, 0xffff0000
; SI-NEXT:    v_writelane_b32 v34, s4, 2
; SI-NEXT:    s_lshl_b32 s4, s16, 16
; SI-NEXT:    s_and_b32 s11, s70, 0xffff0000
; SI-NEXT:    s_lshl_b32 s10, s70, 16
; SI-NEXT:    s_and_b32 s13, s71, 0xffff0000
; SI-NEXT:    s_lshl_b32 s12, s71, 16
; SI-NEXT:    s_and_b32 s15, s80, 0xffff0000
; SI-NEXT:    s_lshl_b32 s14, s80, 16
; SI-NEXT:    s_and_b32 s41, s81, 0xffff0000
; SI-NEXT:    s_lshl_b32 s40, s81, 16
; SI-NEXT:    s_and_b32 s43, s82, 0xffff0000
; SI-NEXT:    s_lshl_b32 s42, s82, 16
; SI-NEXT:    s_and_b32 s45, s83, 0xffff0000
; SI-NEXT:    s_lshl_b32 s44, s83, 16
; SI-NEXT:    s_and_b32 s47, s84, 0xffff0000
; SI-NEXT:    s_lshl_b32 s46, s84, 16
; SI-NEXT:    s_and_b32 s57, s85, 0xffff0000
; SI-NEXT:    s_lshl_b32 s56, s85, 16
; SI-NEXT:    s_and_b32 s59, s86, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s86, 16
; SI-NEXT:    s_and_b32 s61, s87, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s87, 16
; SI-NEXT:    s_and_b32 s63, s96, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s96, 16
; SI-NEXT:    s_and_b32 s73, s97, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s97, 16
; SI-NEXT:    s_and_b32 s75, s98, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s98, 16
; SI-NEXT:    s_and_b32 s77, s99, 0xffff0000
; SI-NEXT:    s_lshl_b32 s76, s99, 16
; SI-NEXT:    s_and_b32 s79, s6, 0xffff0000
; SI-NEXT:    s_lshl_b32 s78, s6, 16
; SI-NEXT:    s_and_b32 s89, s7, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s7, 16
; SI-NEXT:    s_and_b32 s91, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s8, 16
; SI-NEXT:    s_and_b32 s93, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s9, 16
; SI-NEXT:    s_and_b32 s95, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s29, 16
; SI-NEXT:    s_and_b32 s31, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s30, s28, 16
; SI-NEXT:    s_and_b32 s35, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s34, s27, 16
; SI-NEXT:    s_and_b32 s37, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s36, s26, 16
; SI-NEXT:    s_and_b32 s39, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s38, s25, 16
; SI-NEXT:    s_and_b32 s49, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s48, s24, 16
; SI-NEXT:    s_and_b32 s51, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s50, s23, 16
; SI-NEXT:    s_and_b32 s53, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s52, s22, 16
; SI-NEXT:    s_and_b32 s55, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s54, s21, 16
; SI-NEXT:    s_and_b32 s65, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s64, s20, 16
; SI-NEXT:    s_and_b32 s67, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s66, s19, 16
; SI-NEXT:    s_and_b32 s69, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s68, s18, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 3
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_branch .LBB17_3
; SI-NEXT:  .LBB17_2:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr37
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr35
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr31
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr93
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr77
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr75
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr73
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr63
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr61
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr59
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr57
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr47
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr45
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    ; implicit-def: $sgpr43
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr41
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr15
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr13
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr11
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:  .LBB17_3: ; %Flow
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_mov_b32 s4, s10
; SI-NEXT:    s_mov_b32 s5, s11
; SI-NEXT:    s_mov_b32 s10, s12
; SI-NEXT:    s_mov_b32 s11, s13
; SI-NEXT:    s_mov_b32 s12, s14
; SI-NEXT:    s_mov_b32 s13, s15
; SI-NEXT:    s_mov_b32 s14, s40
; SI-NEXT:    s_mov_b32 s15, s41
; SI-NEXT:    s_mov_b32 s40, s42
; SI-NEXT:    s_mov_b32 s41, s43
; SI-NEXT:    s_mov_b32 s42, s44
; SI-NEXT:    s_mov_b32 s43, s45
; SI-NEXT:    s_mov_b32 s44, s46
; SI-NEXT:    s_mov_b32 s45, s47
; SI-NEXT:    s_mov_b32 s46, s56
; SI-NEXT:    s_mov_b32 s47, s57
; SI-NEXT:    s_mov_b32 s56, s58
; SI-NEXT:    s_mov_b32 s57, s59
; SI-NEXT:    s_mov_b32 s58, s60
; SI-NEXT:    s_mov_b32 s59, s61
; SI-NEXT:    s_mov_b32 s60, s62
; SI-NEXT:    s_mov_b32 s61, s63
; SI-NEXT:    s_mov_b32 s62, s72
; SI-NEXT:    s_mov_b32 s63, s73
; SI-NEXT:    s_mov_b32 s72, s74
; SI-NEXT:    s_mov_b32 s73, s75
; SI-NEXT:    s_mov_b32 s74, s76
; SI-NEXT:    s_mov_b32 s75, s77
; SI-NEXT:    s_mov_b32 s76, s78
; SI-NEXT:    s_mov_b32 s77, s79
; SI-NEXT:    v_readlane_b32 s78, v34, 0
; SI-NEXT:    v_readlane_b32 s79, v34, 1
; SI-NEXT:    s_cbranch_vccnz .LBB17_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_add_i32 s6, s6, 3
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s99, s99, 3
; SI-NEXT:    s_add_i32 s98, s98, 3
; SI-NEXT:    s_add_i32 s97, s97, 3
; SI-NEXT:    s_add_i32 s96, s96, 3
; SI-NEXT:    s_add_i32 s87, s87, 3
; SI-NEXT:    s_add_i32 s86, s86, 3
; SI-NEXT:    s_add_i32 s85, s85, 3
; SI-NEXT:    s_add_i32 s84, s84, 3
; SI-NEXT:    s_add_i32 s83, s83, 3
; SI-NEXT:    s_add_i32 s82, s82, 3
; SI-NEXT:    s_add_i32 s81, s81, 3
; SI-NEXT:    s_add_i32 s80, s80, 3
; SI-NEXT:    s_add_i32 s71, s71, 3
; SI-NEXT:    s_add_i32 s70, s70, 3
; SI-NEXT:    s_and_b32 s77, s6, 0xffff0000
; SI-NEXT:    s_lshl_b32 s76, s6, 16
; SI-NEXT:    s_and_b32 s6, s16, 0xffff0000
; SI-NEXT:    s_and_b32 s5, s70, 0xffff0000
; SI-NEXT:    s_lshl_b32 s4, s70, 16
; SI-NEXT:    s_and_b32 s11, s71, 0xffff0000
; SI-NEXT:    s_lshl_b32 s10, s71, 16
; SI-NEXT:    s_and_b32 s13, s80, 0xffff0000
; SI-NEXT:    s_lshl_b32 s12, s80, 16
; SI-NEXT:    s_and_b32 s15, s81, 0xffff0000
; SI-NEXT:    s_lshl_b32 s14, s81, 16
; SI-NEXT:    s_and_b32 s41, s82, 0xffff0000
; SI-NEXT:    s_lshl_b32 s40, s82, 16
; SI-NEXT:    s_and_b32 s43, s83, 0xffff0000
; SI-NEXT:    s_lshl_b32 s42, s83, 16
; SI-NEXT:    s_and_b32 s45, s84, 0xffff0000
; SI-NEXT:    s_lshl_b32 s44, s84, 16
; SI-NEXT:    s_and_b32 s47, s85, 0xffff0000
; SI-NEXT:    s_lshl_b32 s46, s85, 16
; SI-NEXT:    s_and_b32 s57, s86, 0xffff0000
; SI-NEXT:    s_lshl_b32 s56, s86, 16
; SI-NEXT:    s_and_b32 s59, s87, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s87, 16
; SI-NEXT:    s_and_b32 s61, s96, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s96, 16
; SI-NEXT:    s_and_b32 s63, s97, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s97, 16
; SI-NEXT:    s_and_b32 s73, s98, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s98, 16
; SI-NEXT:    s_and_b32 s75, s99, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s99, 16
; SI-NEXT:    s_and_b32 s89, s7, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s7, 16
; SI-NEXT:    s_and_b32 s91, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s8, 16
; SI-NEXT:    s_and_b32 s93, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s9, 16
; SI-NEXT:    s_and_b32 s95, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s29, 16
; SI-NEXT:    s_and_b32 s31, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s30, s28, 16
; SI-NEXT:    s_and_b32 s35, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s34, s27, 16
; SI-NEXT:    s_and_b32 s37, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s36, s26, 16
; SI-NEXT:    s_and_b32 s39, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s38, s25, 16
; SI-NEXT:    s_and_b32 s49, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s48, s24, 16
; SI-NEXT:    s_and_b32 s51, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s50, s23, 16
; SI-NEXT:    s_and_b32 s53, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s52, s22, 16
; SI-NEXT:    s_and_b32 s55, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s54, s21, 16
; SI-NEXT:    s_and_b32 s65, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s64, s20, 16
; SI-NEXT:    s_and_b32 s67, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s66, s19, 16
; SI-NEXT:    s_and_b32 s69, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s68, s18, 16
; SI-NEXT:    s_and_b32 s79, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s78, s17, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v34, s6, 2
; SI-NEXT:    s_lshl_b32 s6, s16, 16
; SI-NEXT:    v_writelane_b32 v34, s6, 3
; SI-NEXT:  .LBB17_5: ; %end
; SI-NEXT:    v_readlane_b32 s6, v34, 2
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 3
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s79
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s78
; SI-NEXT:    v_lshr_b64 v[1:2], v[1:2], 16
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s69
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s68
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s67
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s66
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s65
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s64
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s55
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s54
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s53
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s52
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s51
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s50
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s49
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s48
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s39
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s38
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s37
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s36
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s35
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s34
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s31
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s30
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s95
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s94
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s93
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s92
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s91
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s90
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s89
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s88
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s77
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s76
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s75
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s74
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s73
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s72
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s63
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s62
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s61
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s60
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s59
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s58
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s57
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s56
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s47
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s46
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s45
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s44
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s43
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v26
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s42
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s41
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v27
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s40
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s15
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v28
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s14
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s13
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v29
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s12
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s11
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v30
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s10
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s5
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v31
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s4
; SI-NEXT:    v_lshr_b64 v[31:32], v[31:32], 16
; SI-NEXT:    v_readlane_b32 s99, v33, 35
; SI-NEXT:    v_readlane_b32 s98, v33, 34
; SI-NEXT:    v_readlane_b32 s97, v33, 33
; SI-NEXT:    v_readlane_b32 s96, v33, 32
; SI-NEXT:    v_readlane_b32 s87, v33, 31
; SI-NEXT:    v_readlane_b32 s86, v33, 30
; SI-NEXT:    v_readlane_b32 s85, v33, 29
; SI-NEXT:    v_readlane_b32 s84, v33, 28
; SI-NEXT:    v_readlane_b32 s83, v33, 27
; SI-NEXT:    v_readlane_b32 s82, v33, 26
; SI-NEXT:    v_readlane_b32 s81, v33, 25
; SI-NEXT:    v_readlane_b32 s80, v33, 24
; SI-NEXT:    v_readlane_b32 s71, v33, 23
; SI-NEXT:    v_readlane_b32 s70, v33, 22
; SI-NEXT:    v_readlane_b32 s69, v33, 21
; SI-NEXT:    v_readlane_b32 s68, v33, 20
; SI-NEXT:    v_readlane_b32 s67, v33, 19
; SI-NEXT:    v_readlane_b32 s66, v33, 18
; SI-NEXT:    v_readlane_b32 s65, v33, 17
; SI-NEXT:    v_readlane_b32 s64, v33, 16
; SI-NEXT:    v_readlane_b32 s55, v33, 15
; SI-NEXT:    v_readlane_b32 s54, v33, 14
; SI-NEXT:    v_readlane_b32 s53, v33, 13
; SI-NEXT:    v_readlane_b32 s52, v33, 12
; SI-NEXT:    v_readlane_b32 s51, v33, 11
; SI-NEXT:    v_readlane_b32 s50, v33, 10
; SI-NEXT:    v_readlane_b32 s49, v33, 9
; SI-NEXT:    v_readlane_b32 s48, v33, 8
; SI-NEXT:    v_readlane_b32 s39, v33, 7
; SI-NEXT:    v_readlane_b32 s38, v33, 6
; SI-NEXT:    v_readlane_b32 s37, v33, 5
; SI-NEXT:    v_readlane_b32 s36, v33, 4
; SI-NEXT:    v_readlane_b32 s35, v33, 3
; SI-NEXT:    v_readlane_b32 s34, v33, 2
; SI-NEXT:    v_readlane_b32 s31, v33, 1
; SI-NEXT:    v_readlane_b32 s30, v33, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v64bf16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB17_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB17_3
; VI-NEXT:  .LBB17_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s47, s47, 3
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:  .LBB17_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB17_4:
; VI-NEXT:    s_branch .LBB17_2
;
; GFX9-LABEL: bitcast_v32i32_to_v64bf16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB17_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB17_3
; GFX9-NEXT:  .LBB17_2: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s47, s47, 3
; GFX9-NEXT:    s_add_i32 s29, s29, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s27, s27, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s25, s25, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s23, s23, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s21, s21, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s19, s19, 3
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    s_add_i32 s17, s17, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    s_add_i32 s9, s9, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    s_add_i32 s13, s13, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s15, s15, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s41, s41, 3
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    s_add_i32 s45, s45, 3
; GFX9-NEXT:  .LBB17_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB17_4:
; GFX9-NEXT:    s_branch .LBB17_2
;
; GFX11-LABEL: bitcast_v32i32_to_v64bf16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB17_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB17_3
; GFX11-NEXT:  .LBB17_2: ; %cmp.true
; GFX11-NEXT:    s_add_i32 s27, s27, 3
; GFX11-NEXT:    s_add_i32 s26, s26, 3
; GFX11-NEXT:    s_add_i32 s25, s25, 3
; GFX11-NEXT:    s_add_i32 s24, s24, 3
; GFX11-NEXT:    s_add_i32 s23, s23, 3
; GFX11-NEXT:    s_add_i32 s22, s22, 3
; GFX11-NEXT:    s_add_i32 s21, s21, 3
; GFX11-NEXT:    s_add_i32 s20, s20, 3
; GFX11-NEXT:    s_add_i32 s19, s19, 3
; GFX11-NEXT:    s_add_i32 s18, s18, 3
; GFX11-NEXT:    s_add_i32 s17, s17, 3
; GFX11-NEXT:    s_add_i32 s16, s16, 3
; GFX11-NEXT:    s_add_i32 s3, s3, 3
; GFX11-NEXT:    s_add_i32 s2, s2, 3
; GFX11-NEXT:    s_add_i32 s1, s1, 3
; GFX11-NEXT:    s_add_i32 s0, s0, 3
; GFX11-NEXT:    s_add_i32 s4, s4, 3
; GFX11-NEXT:    s_add_i32 s5, s5, 3
; GFX11-NEXT:    s_add_i32 s6, s6, 3
; GFX11-NEXT:    s_add_i32 s7, s7, 3
; GFX11-NEXT:    s_add_i32 s8, s8, 3
; GFX11-NEXT:    s_add_i32 s9, s9, 3
; GFX11-NEXT:    s_add_i32 s10, s10, 3
; GFX11-NEXT:    s_add_i32 s11, s11, 3
; GFX11-NEXT:    s_add_i32 s12, s12, 3
; GFX11-NEXT:    s_add_i32 s13, s13, 3
; GFX11-NEXT:    s_add_i32 s14, s14, 3
; GFX11-NEXT:    s_add_i32 s15, s15, 3
; GFX11-NEXT:    s_add_i32 s40, s40, 3
; GFX11-NEXT:    s_add_i32 s41, s41, 3
; GFX11-NEXT:    s_add_i32 s29, s29, 3
; GFX11-NEXT:    s_add_i32 s28, s28, 3
; GFX11-NEXT:  .LBB17_3: ; %end
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB17_4:
; GFX11-NEXT:    s_branch .LBB17_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define <32 x i32> @bitcast_v64bf16_to_v32i32(<64 x bfloat> %a, i32 %b) {
; SI-LABEL: bitcast_v64bf16_to_v32i32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v0
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v63
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v62
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v60
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v58, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_and_b32_e32 v46, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v36, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v53
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v37
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v38
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v38
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v42
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v40
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v55
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v18
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v19
; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v51
; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v20
; SI-NEXT:    v_mul_f32_e32 v45, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v49
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v41, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v23
; SI-NEXT:    v_mul_f32_e32 v55, 1.0, v39
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v53, 1.0, v36
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v25
; SI-NEXT:    v_mul_f32_e32 v51, 1.0, v35
; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v49, 1.0, v34
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v39, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v29
; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v31
; SI-NEXT:    v_mul_f32_e32 v35, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v33, 1.0, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB18_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v51
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v39
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v36
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v34
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_alignbit_b32 v19, v19, v59, 16
; SI-NEXT:    v_alignbit_b32 v20, v20, v57, 16
; SI-NEXT:    v_alignbit_b32 v21, v21, v46, 16
; SI-NEXT:    v_alignbit_b32 v22, v22, v44, 16
; SI-NEXT:    v_alignbit_b32 v23, v23, v42, 16
; SI-NEXT:    v_alignbit_b32 v24, v24, v40, 16
; SI-NEXT:    v_alignbit_b32 v25, v25, v54, 16
; SI-NEXT:    v_alignbit_b32 v26, v26, v52, 16
; SI-NEXT:    v_alignbit_b32 v27, v27, v50, 16
; SI-NEXT:    v_alignbit_b32 v28, v28, v48, 16
; SI-NEXT:    v_alignbit_b32 v29, v29, v38, 16
; SI-NEXT:    v_alignbit_b32 v30, v30, v35, 16
; SI-NEXT:    v_alignbit_b32 v31, v31, v33, 16
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_alignbit_b32 v0, v0, v63, 16
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v62
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v60
; SI-NEXT:    v_alignbit_b32 v18, v18, v61, 16
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:  .LBB18_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB18_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v63
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v60
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v55
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v0, v1, v0, 16
; SI-NEXT:    v_alignbit_b32 v1, v3, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v2, v3, v2, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_alignbit_b32 v3, v4, v3, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_alignbit_b32 v4, v5, v4, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_alignbit_b32 v5, v6, v5, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_alignbit_b32 v6, v7, v6, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_alignbit_b32 v7, v8, v7, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_alignbit_b32 v8, v9, v8, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_alignbit_b32 v9, v10, v9, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_alignbit_b32 v10, v11, v10, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_alignbit_b32 v11, v12, v11, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_alignbit_b32 v12, v13, v12, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v13, v14, v13, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_alignbit_b32 v14, v15, v14, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v15, v16, v15, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v16, v17, v16, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_alignbit_b32 v17, v18, v17, 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v61
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_alignbit_b32 v18, v19, v18, 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v59
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_alignbit_b32 v19, v20, v19, 16
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v57
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_alignbit_b32 v20, v21, v20, 16
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_alignbit_b32 v21, v22, v21, 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v44
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_alignbit_b32 v22, v23, v22, 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_alignbit_b32 v23, v24, v23, 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_alignbit_b32 v24, v25, v24, 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_alignbit_b32 v25, v26, v25, 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_alignbit_b32 v26, v27, v26, 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_alignbit_b32 v27, v28, v27, 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_alignbit_b32 v28, v29, v28, 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_alignbit_b32 v29, v30, v29, 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v35
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_alignbit_b32 v30, v31, v30, 16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v33
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_alignbit_b32 v31, v32, v31, 16
; SI-NEXT:  .LBB18_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v32i32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB18_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v15, 16, 1
; VI-NEXT:    s_movk_i32 s6, 0x7fff
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v15
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_alignbit_b32 v15, v15, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v14, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v14
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; VI-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_alignbit_b32 v14, v14, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v13
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_alignbit_b32 v13, v13, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v12, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v12
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; VI-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_alignbit_b32 v12, v12, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v11
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_alignbit_b32 v11, v11, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v10, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v10
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; VI-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_alignbit_b32 v10, v10, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v9
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_alignbit_b32 v9, v9, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v8, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v8
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; VI-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_alignbit_b32 v8, v8, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v7
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_alignbit_b32 v7, v7, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v6, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v6
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; VI-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_alignbit_b32 v6, v6, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v5
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_alignbit_b32 v5, v5, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v4, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v4
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; VI-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_alignbit_b32 v4, v4, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v3
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_alignbit_b32 v3, v3, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v2
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_alignbit_b32 v2, v2, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v1
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_alignbit_b32 v1, v1, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v0
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; VI-NEXT:    v_alignbit_b32 v0, v0, v32, 16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v30, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v30
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; VI-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_alignbit_b32 v30, v30, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v29, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v29
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_alignbit_b32 v29, v29, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v28, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v28
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; VI-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_alignbit_b32 v28, v28, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v27
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_alignbit_b32 v27, v27, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v26, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v26
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; VI-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_alignbit_b32 v26, v26, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v25
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_alignbit_b32 v25, v25, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v24, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v24
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; VI-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_alignbit_b32 v24, v24, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v23
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_alignbit_b32 v23, v23, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v22, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v22
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; VI-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_alignbit_b32 v22, v22, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v21
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_alignbit_b32 v21, v21, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v20, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v20
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; VI-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_alignbit_b32 v20, v20, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v19
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_alignbit_b32 v19, v19, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v18
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_alignbit_b32 v18, v18, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v17
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_alignbit_b32 v17, v17, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v16
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_alignbit_b32 v16, v16, v32, 16
; VI-NEXT:  .LBB18_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v32i32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB18_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    s_movk_i32 s6, 0x7fff
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v15, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v15, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; GFX9-NEXT:    s_mov_b32 s7, 0x7060302
; GFX9-NEXT:    v_perm_b32 v15, v15, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v14, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v14, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v14, v14, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v13, v13, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v12, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v12, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v12, v12, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v11, v11, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v10, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v10, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v10, v10, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v9, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v9, v9, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v8, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v8, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v8, v8, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v7, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v7, v7, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v6, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v6, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v6, v6, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v5, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v5, v5, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v4, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v4, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v4, v4, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v3, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v3, v3, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v2, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v2, v2, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v1, v1, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v0, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v0, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v0, v0, v32, s7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v30, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v30, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v30, v30, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v29, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v29, v29, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v28, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v28, v28, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v27, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v27, v27, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v26, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v26, v26, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v25, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v25, v25, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v24, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v24, v24, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v23, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v23, v23, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v22, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v22, v22, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v21, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v21, v21, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v20, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v20, v20, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v19, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v19, v19, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v18, v18, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v17, v17, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v16, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v16, v16, v32, s7
; GFX9-NEXT:  .LBB18_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v32i32:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB18_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v63, v31 :: v_dual_mov_b32 v62, v30
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v61, v29 :: v_dual_mov_b32 v60, v28
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v59, v27 :: v_dual_mov_b32 v58, v26
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v57, v25 :: v_dual_mov_b32 v56, v24
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v23 :: v_dual_mov_b32 v54, v22
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v21 :: v_dual_mov_b32 v52, v20
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v19 :: v_dual_mov_b32 v50, v18
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v17 :: v_dual_mov_b32 v48, v16
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v47, v15 :: v_dual_mov_b32 v46, v14
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v45, v13 :: v_dual_mov_b32 v44, v12
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v43, v11 :: v_dual_mov_b32 v42, v10
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v41, v9 :: v_dual_mov_b32 v40, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v7 :: v_dual_mov_b32 v38, v6
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v5 :: v_dual_mov_b32 v36, v4
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v3 :: v_dual_mov_b32 v34, v2
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v1 :: v_dual_mov_b32 v32, v0
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT:  .LBB18_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB18_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v32, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v35, v38 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v39, v33, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v47, v36, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v32.h
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_add3_u32 v38, v48, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v38, v49, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v34 :: v_dual_cndmask_b32 v46, v35, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, v14.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v33, v34, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v13, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v45, v33, v35
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v13.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v15, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v44, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, v12.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v11, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v14, v32 :: v_dual_and_b32 v12, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v43, v15, v33
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v10, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v13, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v42, v14, v32, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v10.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v9, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v41, v13, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, v9.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v8, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v11, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v40, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v8.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v7, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v39, v11, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v7.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v9, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v6.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v8, v10 :: v_dual_and_b32 v6, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v37, v9, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v8, v10, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v4.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v6, v8 :: v_dual_and_b32 v4, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_cndmask_b32 v35, v7, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v5, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v6, v8, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v4, v6 :: v_dual_and_b32 v2, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_cndmask_b32 v33, v5, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v1.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v31
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v5, 16, v31
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v3, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v4, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v0.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v63, v3, v7 :: v_dual_and_b32 v0, 0xffff0000, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v62, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v28
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v61, v3, v7 :: v_dual_lshlrev_b32 v4, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v60, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v59, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v25
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v4, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v58, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v24
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v57, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v56, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v55, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v5, 16, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v53, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v52, v0, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v2.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_lshlrev_b32 v6, 16, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v6 :: v_dual_cndmask_b32 v51, v0, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v8, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v50, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v5, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v16
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v1 :: v_dual_add_f32 v1, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v0.h
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v9, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v1, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v5.h
; GFX11-TRUE16-NEXT:  .LBB18_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v40 :: v_dual_mov_b32 v9, v41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v42 :: v_dual_mov_b32 v11, v43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, v44 :: v_dual_mov_b32 v13, v45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v46 :: v_dual_mov_b32 v15, v47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v56 :: v_dual_mov_b32 v25, v57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v58 :: v_dual_mov_b32 v27, v59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v60 :: v_dual_mov_b32 v29, v61
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v62 :: v_dual_mov_b32 v31, v63
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v33
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v34 :: v_dual_mov_b32 v3, v35
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v36 :: v_dual_mov_b32 v5, v37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v38 :: v_dual_mov_b32 v7, v39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v48 :: v_dual_mov_b32 v17, v49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v50 :: v_dual_mov_b32 v19, v51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v52 :: v_dual_mov_b32 v21, v53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v54 :: v_dual_mov_b32 v23, v55
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v32i32:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB18_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v39, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v13, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v11, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v8, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v9, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v34 :: v_dual_add_f32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v6, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v6, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v7, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v5, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v3, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v0, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v1, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v31
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v31, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v29
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v30, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v29, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v28, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v27
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v28, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v27, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v26, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v25, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v23
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v23, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v36, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v21, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v32, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v38, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v20, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v19, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v48, v18, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v36, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v36
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v39, v17, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v16, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v35, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v37, v50, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v16, v36, 0x7060302
; GFX11-FAKE16-NEXT:  .LBB18_2: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define inreg <32 x i32> @bitcast_v64bf16_to_v32i32_scalar(<64 x bfloat> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64bf16_to_v32i32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    s_and_b32 s12, s25, 0xffff0000
; SI-NEXT:    s_and_b32 s30, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v1
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; SI-NEXT:    s_and_b32 s34, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_readfirstlane_b32 s94, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s34
; SI-NEXT:    v_readfirstlane_b32 s90, v5
; SI-NEXT:    s_and_b32 s95, s94, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s30
; SI-NEXT:    v_readfirstlane_b32 s88, v6
; SI-NEXT:    s_and_b32 s91, s90, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s95
; SI-NEXT:    v_readfirstlane_b32 s78, v7
; SI-NEXT:    s_and_b32 s89, s88, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s91
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    s_and_b32 s79, s78, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s89
; SI-NEXT:    v_readfirstlane_b32 s74, v9
; SI-NEXT:    s_and_b32 s77, s76, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s79
; SI-NEXT:    v_readfirstlane_b32 s72, v10
; SI-NEXT:    s_and_b32 s75, s74, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s77
; SI-NEXT:    v_readfirstlane_b32 s62, v11
; SI-NEXT:    s_and_b32 s73, s72, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s75
; SI-NEXT:    v_readfirstlane_b32 s60, v12
; SI-NEXT:    s_and_b32 s63, s62, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s73
; SI-NEXT:    v_readfirstlane_b32 s58, v13
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s63
; SI-NEXT:    v_readfirstlane_b32 s56, v14
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s61
; SI-NEXT:    v_readfirstlane_b32 s46, v15
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s59
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s57
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s47
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    v_readfirstlane_b32 s92, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s45
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s29, 16
; SI-NEXT:    s_lshl_b32 s7, s28, 16
; SI-NEXT:    s_and_b32 s8, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s9, s27, 16
; SI-NEXT:    s_and_b32 s10, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s11, s26, 16
; SI-NEXT:    s_lshl_b32 s13, s25, 16
; SI-NEXT:    s_and_b32 s14, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s15, s24, 16
; SI-NEXT:    s_and_b32 s24, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s25, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s26, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s27, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s28, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s29, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_and_b32 s41, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_and_b32 s93, s92, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_and_b32 s36, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s43
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cmp_lg_u32 vcc_lo, 0
; SI-NEXT:    v_mul_f32_e64 v55, 1.0, s41
; SI-NEXT:    v_mul_f32_e64 v53, 1.0, s40
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s29
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v62, 1.0, s27
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s26
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s25
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s24
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s14
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s10
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s8
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s4
; SI-NEXT:    v_mul_f32_e64 v39, 1.0, s36
; SI-NEXT:    v_mul_f32_e64 v47, 1.0, s93
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e64 v43, 1.0, s16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s17
; SI-NEXT:    v_mul_f32_e64 v34, 1.0, s18
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s19
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s20
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s21
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s22
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s15
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s13
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s11
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s9
; SI-NEXT:    v_mul_f32_e64 v61, 1.0, s7
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s5
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s37
; SI-NEXT:    v_mul_f32_e64 v60, 1.0, s35
; SI-NEXT:    v_mul_f32_e64 v33, 1.0, s31
; SI-NEXT:    v_mul_f32_e64 v57, 1.0, s94
; SI-NEXT:    v_mul_f32_e64 v59, 1.0, s92
; SI-NEXT:    v_mul_f32_e64 v56, 1.0, s90
; SI-NEXT:    v_mul_f32_e64 v38, 1.0, s88
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s78
; SI-NEXT:    v_mul_f32_e64 v36, 1.0, s76
; SI-NEXT:    v_mul_f32_e64 v46, 1.0, s74
; SI-NEXT:    v_mul_f32_e64 v48, 1.0, s72
; SI-NEXT:    v_mul_f32_e64 v49, 1.0, s62
; SI-NEXT:    v_mul_f32_e64 v45, 1.0, s60
; SI-NEXT:    v_mul_f32_e64 v42, 1.0, s58
; SI-NEXT:    v_mul_f32_e64 v51, 1.0, s56
; SI-NEXT:    v_mul_f32_e64 v41, 1.0, s46
; SI-NEXT:    v_mul_f32_e64 v54, 1.0, s44
; SI-NEXT:    v_mul_f32_e64 v52, 1.0, s42
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB19_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v55
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v53
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v44, v1
; SI-NEXT:    v_mov_b32_e32 v43, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v12
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    v_mov_b32_e32 v36, v47
; SI-NEXT:    v_mov_b32_e32 v50, v49
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v40, v41
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v1, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[43:44], 16
; SI-NEXT:    v_mov_b32_e32 v2, v34
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[2:3], v[34:35], 16
; SI-NEXT:    v_mov_b32_e32 v3, v17
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[3:4], v[17:18], 16
; SI-NEXT:    v_mov_b32_e32 v4, v20
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[4:5], v[20:21], 16
; SI-NEXT:    v_mov_b32_e32 v5, v19
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v6
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[5:6], v[19:20], 16
; SI-NEXT:    v_mov_b32_e32 v6, v13
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[6:7], v[13:14], 16
; SI-NEXT:    v_mov_b32_e32 v7, v31
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[7:8], v[31:32], 16
; SI-NEXT:    v_mov_b32_e32 v8, v23
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[8:9], v[23:24], 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v11
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v44, v45
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v9
; SI-NEXT:    v_mov_b32_e32 v9, v25
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[9:10], v[25:26], 16
; SI-NEXT:    v_mov_b32_e32 v10, v30
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v11, v27
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[11:12], v[27:28], 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v18
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v12
; SI-NEXT:    v_mov_b32_e32 v12, v61
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[12:13], v[61:62], 16
; SI-NEXT:    v_mov_b32_e32 v13, v29
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[13:14], v[29:30], 16
; SI-NEXT:    v_mov_b32_e32 v14, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v62, v39
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[14:15], v[15:16], 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v15
; SI-NEXT:    v_mov_b32_e32 v15, v60
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[15:16], v[60:61], 16
; SI-NEXT:    v_mov_b32_e32 v16, v33
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[16:17], v[33:34], 16
; SI-NEXT:    v_mov_b32_e32 v17, v57
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[17:18], v[57:58], 16
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v47
; SI-NEXT:    v_lshr_b64 v[18:19], v[59:60], 16
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v20
; SI-NEXT:    v_lshr_b64 v[19:20], v[56:57], 16
; SI-NEXT:    v_mov_b32_e32 v20, v38
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[20:21], v[38:39], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; SI-NEXT:    v_mov_b32_e32 v39, v62
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v21
; SI-NEXT:    v_mov_b32_e32 v21, v22
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[21:22], v[22:23], 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v55
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[37:38], 16
; SI-NEXT:    v_lshr_b64 v[23:24], v[46:47], 16
; SI-NEXT:    v_lshr_b64 v[24:25], v[48:49], 16
; SI-NEXT:    v_lshr_b64 v[25:26], v[50:51], 16
; SI-NEXT:    v_lshr_b64 v[26:27], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[27:28], v[42:43], 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_lshr_b64 v[61:62], v[52:53], 16
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v45, v44
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v47, v36
; SI-NEXT:    v_mov_b32_e32 v36, v37
; SI-NEXT:    v_mov_b32_e32 v49, v50
; SI-NEXT:    v_mov_b32_e32 v51, v33
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_lshr_b64 v[29:30], v[40:41], 16
; SI-NEXT:    v_lshr_b64 v[30:31], v[54:55], 16
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v31, v61
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    s_branch .LBB19_3
; SI-NEXT:  .LBB19_2:
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB19_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v37, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v33, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(3)
; SI-NEXT:    v_mov_b32_e32 v38, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v50, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v40, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_cbranch_vccnz .LBB19_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v55
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v34, v56
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshr_b64 v[1:2], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v33
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_lshr_b64 v[32:33], v[32:33], 16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v37
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v31, v32
; SI-NEXT:  .LBB19_5: ; %end
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v32i32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v35, s36, 0
; VI-NEXT:    v_writelane_b32 v35, s37, 1
; VI-NEXT:    v_writelane_b32 v35, s38, 2
; VI-NEXT:    v_writelane_b32 v35, s39, 3
; VI-NEXT:    v_writelane_b32 v35, s48, 4
; VI-NEXT:    v_writelane_b32 v35, s49, 5
; VI-NEXT:    v_writelane_b32 v35, s50, 6
; VI-NEXT:    v_writelane_b32 v35, s51, 7
; VI-NEXT:    v_writelane_b32 v35, s52, 8
; VI-NEXT:    v_writelane_b32 v35, s53, 9
; VI-NEXT:    v_writelane_b32 v35, s54, 10
; VI-NEXT:    v_writelane_b32 v35, s55, 11
; VI-NEXT:    v_writelane_b32 v35, s64, 12
; VI-NEXT:    v_writelane_b32 v35, s65, 13
; VI-NEXT:    v_writelane_b32 v35, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v35, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB19_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB19_4
; VI-NEXT:  .LBB19_2: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x40c00000
; VI-NEXT:    s_lshl_b32 s4, s50, 16
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    s_and_b32 s4, s50, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v2, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s48, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; VI-NEXT:    v_add_f32_e32 v2, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    s_and_b32 s4, s48, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v4, v3, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_lshrrev_b64 v[14:15], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s46, 16
; VI-NEXT:    v_lshrrev_b64 v[12:13], 16, v[2:3]
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[10:11], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[8:9], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s38, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[4:5], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s36, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[2:3], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; VI-NEXT:    v_bfe_u32 v3, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v0
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_lshrrev_b64 v[0:1], 16, v[0:1]
; VI-NEXT:    s_and_b32 s4, s66, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s66, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s64, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s64, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s62, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[30:31], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s62, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s60, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[28:29], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s60, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s58, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[26:27], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s58, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s56, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[24:25], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s56, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s54, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[22:23], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s54, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s52, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[20:21], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s52, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s37, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s37, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_and_b32 s4, s39, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v5, v7, vcc
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    s_lshl_b32 s4, s39, 16
; VI-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_mov_b32_e32 v1, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v7, v9, vcc
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    s_lshl_b32 s4, s41, 16
; VI-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v3
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_mov_b32_e32 v3, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v9, v11, vcc
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    s_lshl_b32 s4, s43, 16
; VI-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_mov_b32_e32 v5, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v11, v13, vcc
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    s_lshl_b32 s4, s45, 16
; VI-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[16:17], 16, v[15:16]
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_mov_b32_e32 v7, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v13, v15, vcc
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    s_lshl_b32 s4, s47, 16
; VI-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_and_b32 s4, s49, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_mov_b32_e32 v9, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v15, v17, vcc
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    s_lshl_b32 s4, s49, 16
; VI-NEXT:    v_cndmask_b32_e32 v13, v15, v17, vcc
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_and_b32 s4, s51, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_mov_b32_e32 v11, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v17, v19, vcc
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    s_lshl_b32 s4, s51, 16
; VI-NEXT:    v_cndmask_b32_e32 v15, v17, v19, vcc
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v13
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    s_and_b32 s4, s53, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_mov_b32_e32 v13, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v19, v21, vcc
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    s_lshl_b32 s4, s53, 16
; VI-NEXT:    v_cndmask_b32_e32 v17, v19, v21, vcc
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    s_and_b32 s4, s55, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_mov_b32_e32 v15, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v21, v23, vcc
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    s_lshl_b32 s4, s55, 16
; VI-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    s_and_b32 s4, s57, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_mov_b32_e32 v17, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v23, v25, vcc
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    s_lshl_b32 s4, s57, 16
; VI-NEXT:    v_cndmask_b32_e32 v21, v23, v25, vcc
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    s_and_b32 s4, s59, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_mov_b32_e32 v19, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v25, v27, vcc
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    s_lshl_b32 s4, s59, 16
; VI-NEXT:    v_cndmask_b32_e32 v23, v25, v27, vcc
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v21
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    s_and_b32 s4, s61, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_mov_b32_e32 v21, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v27, v29, vcc
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    s_lshl_b32 s4, s61, 16
; VI-NEXT:    v_cndmask_b32_e32 v25, v27, v29, vcc
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    s_and_b32 s4, s63, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_mov_b32_e32 v23, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v29, v31, vcc
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    s_lshl_b32 s4, s63, 16
; VI-NEXT:    v_cndmask_b32_e32 v27, v29, v31, vcc
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    s_and_b32 s4, s65, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v25, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v31, v33, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    v_mov_b32_e32 v27, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    s_lshl_b32 s4, s65, 16
; VI-NEXT:    v_cndmask_b32_e32 v29, v31, v33, vcc
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    s_and_b32 s4, s67, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_mov_b32_e32 v29, v33
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    s_lshl_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v32, v31, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_add_u32_e32 v32, vcc, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x7fff, v32
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v32, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; VI-NEXT:    v_lshrrev_b64 v[31:32], 16, v[31:32]
; VI-NEXT:    s_branch .LBB19_5
; VI-NEXT:  .LBB19_3:
; VI-NEXT:    s_branch .LBB19_2
; VI-NEXT:  .LBB19_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB19_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v35, 15
; VI-NEXT:    v_readlane_b32 s66, v35, 14
; VI-NEXT:    v_readlane_b32 s65, v35, 13
; VI-NEXT:    v_readlane_b32 s64, v35, 12
; VI-NEXT:    v_readlane_b32 s55, v35, 11
; VI-NEXT:    v_readlane_b32 s54, v35, 10
; VI-NEXT:    v_readlane_b32 s53, v35, 9
; VI-NEXT:    v_readlane_b32 s52, v35, 8
; VI-NEXT:    v_readlane_b32 s51, v35, 7
; VI-NEXT:    v_readlane_b32 s50, v35, 6
; VI-NEXT:    v_readlane_b32 s49, v35, 5
; VI-NEXT:    v_readlane_b32 s48, v35, 4
; VI-NEXT:    v_readlane_b32 s39, v35, 3
; VI-NEXT:    v_readlane_b32 s38, v35, 2
; VI-NEXT:    v_readlane_b32 s37, v35, 1
; VI-NEXT:    v_readlane_b32 s36, v35, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v32i32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v36, s36, 0
; GFX9-NEXT:    v_writelane_b32 v36, s37, 1
; GFX9-NEXT:    v_writelane_b32 v36, s38, 2
; GFX9-NEXT:    v_writelane_b32 v36, s39, 3
; GFX9-NEXT:    v_writelane_b32 v36, s48, 4
; GFX9-NEXT:    v_writelane_b32 v36, s49, 5
; GFX9-NEXT:    v_writelane_b32 v36, s50, 6
; GFX9-NEXT:    v_writelane_b32 v36, s51, 7
; GFX9-NEXT:    v_writelane_b32 v36, s52, 8
; GFX9-NEXT:    v_writelane_b32 v36, s53, 9
; GFX9-NEXT:    v_writelane_b32 v36, s54, 10
; GFX9-NEXT:    v_writelane_b32 v36, s55, 11
; GFX9-NEXT:    v_writelane_b32 v36, s64, 12
; GFX9-NEXT:    v_writelane_b32 v36, s65, 13
; GFX9-NEXT:    v_writelane_b32 v36, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v36, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB19_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB19_4
; GFX9-NEXT:  .LBB19_2: ; %cmp.true
; GFX9-NEXT:    s_and_b32 s4, s51, 0xffff0000
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x40c00000
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s51, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    s_and_b32 s4, s50, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_add_f32_e32 v2, s4, v16
; GFX9-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v3, v3, v2
; GFX9-NEXT:    v_add_u32_e32 v3, 0x7fff, v3
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    s_lshl_b32 s4, s50, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT:    v_add_f32_e32 v3, s4, v16
; GFX9-NEXT:    v_mov_b32_e32 v32, 0xffff
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    s_and_b32 s4, s49, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_lshl_or_b32 v15, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_and_b32_sdwa v3, v32, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_lshl_or_b32 v14, v2, 16, v3
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s49, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s48, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s48, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s47, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s47, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s46, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s46, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s45, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s45, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s44, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s44, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s43, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s43, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s42, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s42, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s41, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s41, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s40, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s40, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s39, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s39, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s38, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s38, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v17, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s37, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s37, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v17, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v1
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v17, v18, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s36, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v0
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s36, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v17, v18, vcc
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v17, v32, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s67, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s67, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s66, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v31, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s66, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s65, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v30, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s65, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s64, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v29, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s64, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s63, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v28, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s63, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s62, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v27, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s62, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s61, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v26, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s61, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s60, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s60, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s59, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v24, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s59, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s58, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v23, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s58, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s57, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v22, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s57, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s56, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v21, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s56, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s55, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v20, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s55, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v33, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s54, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v19, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s54, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v33, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v18
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s53, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v17
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s53, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v33, v32, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s52, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v33
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    s_lshl_b32 s4, s52, 16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v16, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v16, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v34, v35, vcc
; GFX9-NEXT:    v_and_b32_sdwa v16, v32, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v16, v32, 16, v16
; GFX9-NEXT:    s_branch .LBB19_5
; GFX9-NEXT:  .LBB19_3:
; GFX9-NEXT:    s_branch .LBB19_2
; GFX9-NEXT:  .LBB19_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB19_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v36, 15
; GFX9-NEXT:    v_readlane_b32 s66, v36, 14
; GFX9-NEXT:    v_readlane_b32 s65, v36, 13
; GFX9-NEXT:    v_readlane_b32 s64, v36, 12
; GFX9-NEXT:    v_readlane_b32 s55, v36, 11
; GFX9-NEXT:    v_readlane_b32 s54, v36, 10
; GFX9-NEXT:    v_readlane_b32 s53, v36, 9
; GFX9-NEXT:    v_readlane_b32 s52, v36, 8
; GFX9-NEXT:    v_readlane_b32 s51, v36, 7
; GFX9-NEXT:    v_readlane_b32 s50, v36, 6
; GFX9-NEXT:    v_readlane_b32 s49, v36, 5
; GFX9-NEXT:    v_readlane_b32 s48, v36, 4
; GFX9-NEXT:    v_readlane_b32 s39, v36, 3
; GFX9-NEXT:    v_readlane_b32 s38, v36, 2
; GFX9-NEXT:    v_readlane_b32 s37, v36, 1
; GFX9-NEXT:    v_readlane_b32 s36, v36, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v32i32_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s36, s0
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-TRUE16-NEXT:    s_mov_b32 s48, s24
; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
; GFX11-TRUE16-NEXT:    s_mov_b32 s39, s3
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-TRUE16-NEXT:    s_mov_b32 s49, s25
; GFX11-TRUE16-NEXT:    s_mov_b32 s38, s2
; GFX11-TRUE16-NEXT:    s_mov_b32 s37, s1
; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-TRUE16-NEXT:    s_mov_b32 s50, s26
; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-TRUE16-NEXT:    s_mov_b32 s51, s27
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-TRUE16-NEXT:    s_mov_b32 s52, s28
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-TRUE16-NEXT:    s_mov_b32 s53, s29
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB19_3
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB19_4
; GFX11-TRUE16-NEXT:  .LBB19_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s51, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s50, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s57, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v0.l
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v2.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v4, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v7, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v16, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v18, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v17 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v16, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v16 :: v_dual_add_nc_u32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v17, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v21, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s37, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v0.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v20 :: v_dual_add_nc_u32 v0, v16, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s36, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v22, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v16.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s67, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v24, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s66, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s65, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v20, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_nc_u32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s64, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s63, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s62, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s61, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v16.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s60, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s58, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v33, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s57, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v33 :: v_dual_add_nc_u32 v19, v19, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v32
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v33, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v20, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s55, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v20, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v35, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s52, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v19, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v33, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v19, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, v35, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v34, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v17.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v36, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v16, v34
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v36, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v35 :: v_dual_add_nc_u32 v35, v37, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v33
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v34, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v37, v38, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v17.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v35, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v32
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v16.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v33.l
; GFX11-TRUE16-NEXT:    s_branch .LBB19_5
; GFX11-TRUE16-NEXT:  .LBB19_3:
; GFX11-TRUE16-NEXT:    s_branch .LBB19_2
; GFX11-TRUE16-NEXT:  .LBB19_4:
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-TRUE16-NEXT:  .LBB19_5: ; %end
; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v32i32_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s36, s0
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-FAKE16-NEXT:    s_mov_b32 s47, s23
; GFX11-FAKE16-NEXT:    s_mov_b32 s46, s22
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-FAKE16-NEXT:    s_mov_b32 s45, s21
; GFX11-FAKE16-NEXT:    s_mov_b32 s44, s20
; GFX11-FAKE16-NEXT:    s_mov_b32 s43, s19
; GFX11-FAKE16-NEXT:    s_mov_b32 s42, s18
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-FAKE16-NEXT:    s_mov_b32 s48, s24
; GFX11-FAKE16-NEXT:    s_mov_b32 s41, s17
; GFX11-FAKE16-NEXT:    s_mov_b32 s40, s16
; GFX11-FAKE16-NEXT:    s_mov_b32 s39, s3
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-FAKE16-NEXT:    s_mov_b32 s49, s25
; GFX11-FAKE16-NEXT:    s_mov_b32 s38, s2
; GFX11-FAKE16-NEXT:    s_mov_b32 s37, s1
; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-FAKE16-NEXT:    s_mov_b32 s50, s26
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-FAKE16-NEXT:    s_mov_b32 s51, s27
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-FAKE16-NEXT:    s_mov_b32 s52, s28
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-FAKE16-NEXT:    s_mov_b32 s53, s29
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB19_3
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB19_4
; GFX11-FAKE16-NEXT:  .LBB19_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s51, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s50, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s49, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s49, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s48, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v7 :: v_dual_add_nc_u32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s47, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s46, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s45, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s44, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s43, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s42, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s41, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s40, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v17 :: v_dual_add_nc_u32 v5, v7, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s39, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s38, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s37, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v4, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s67, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v2, v22 :: v_dual_add_nc_u32 v17, v17, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v22, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s66, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s65, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v22, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s64, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s63, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s62, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s61, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s60, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s58, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v32, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v22, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v19, v21 :: v_dual_and_b32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s57, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v18, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s57, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v22, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v23, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v33, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v35, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v21, v33, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v33, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v22, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s55, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v34, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v33, v34 :: v_dual_add_nc_u32 v33, v35, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v37, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v33, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s52, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v38, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v35, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, v48, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v39
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v48
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v39, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v36, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v37, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v17, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v34, 16, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v32, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v36, 16, v33
; GFX11-FAKE16-NEXT:    s_branch .LBB19_5
; GFX11-FAKE16-NEXT:  .LBB19_3:
; GFX11-FAKE16-NEXT:    s_branch .LBB19_2
; GFX11-FAKE16-NEXT:  .LBB19_4:
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-FAKE16-NEXT:  .LBB19_5: ; %end
; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define <64 x half> @bitcast_v32i32_to_v64f16(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v64f16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB20_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB20_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB20_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB20_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
; SI-NEXT:    v_or_b32_e32 v0, v0, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v44
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v42
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v61
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v47
; SI-NEXT:    v_or_b32_e32 v3, v3, v44
; SI-NEXT:    v_or_b32_e32 v5, v5, v42
; SI-NEXT:    v_or_b32_e32 v6, v6, v55
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v60
; SI-NEXT:    v_or_b32_e32 v8, v8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v57
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v45
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v41
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v54
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v52
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v53
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v64f16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB20_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:  .LBB20_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32i32_to_v64f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB20_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11
; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10
; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9
; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8
; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7
; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6
; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5
; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4
; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3
; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16
; GFX9-NEXT:  .LBB20_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32i32_to_v64f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB20_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-NEXT:  .LBB20_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define inreg <64 x half> @bitcast_v32i32_to_v64f16_scalar(<32 x i32> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32i32_to_v64f16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_readfirstlane_b32 s44, v18
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_readfirstlane_b32 s7, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    v_readfirstlane_b32 s9, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_readfirstlane_b32 s10, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s15, v7
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s40, v4
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    v_readfirstlane_b32 s42, v2
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    s_cbranch_scc0 .LBB21_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB21_3
; SI-NEXT:  .LBB21_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s44, s44, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s42, s42, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s40, s40, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s10, s10, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s6, s6, 3
; SI-NEXT:    s_add_i32 s5, s5, 3
; SI-NEXT:    s_add_i32 s4, s4, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:  .LBB21_3: ; %end
; SI-NEXT:    s_lshl_b32 s47, s36, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s69, 16
; SI-NEXT:    s_or_b32 s17, s17, s47
; SI-NEXT:    s_lshl_b32 s47, s34, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s47
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s68, 16
; SI-NEXT:    s_or_b32 s19, s19, s47
; SI-NEXT:    s_lshl_b32 s47, s30, 16
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s47
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s67, 16
; SI-NEXT:    s_or_b32 s21, s21, s47
; SI-NEXT:    s_lshl_b32 s47, s94, 16
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s47
; SI-NEXT:    s_and_b32 s23, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s66, 16
; SI-NEXT:    s_or_b32 s23, s23, s47
; SI-NEXT:    s_lshl_b32 s47, s92, 16
; SI-NEXT:    s_and_b32 s24, s24, 0xffff
; SI-NEXT:    s_or_b32 s24, s24, s47
; SI-NEXT:    s_and_b32 s25, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s65, 16
; SI-NEXT:    s_or_b32 s25, s25, s47
; SI-NEXT:    s_and_b32 s26, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s90, 16
; SI-NEXT:    s_or_b32 s26, s26, s47
; SI-NEXT:    s_and_b32 s27, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s64, 16
; SI-NEXT:    s_or_b32 s27, s27, s47
; SI-NEXT:    s_and_b32 s28, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s88, 16
; SI-NEXT:    s_or_b32 s28, s28, s47
; SI-NEXT:    s_and_b32 s29, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s55, 16
; SI-NEXT:    s_or_b32 s29, s29, s47
; SI-NEXT:    s_and_b32 s44, s44, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s78, 16
; SI-NEXT:    s_or_b32 s44, s44, s47
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s54, 16
; SI-NEXT:    s_or_b32 s45, s45, s47
; SI-NEXT:    s_and_b32 s42, s42, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s76, 16
; SI-NEXT:    s_or_b32 s42, s42, s47
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s53, 16
; SI-NEXT:    s_or_b32 s43, s43, s47
; SI-NEXT:    s_and_b32 s40, s40, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s74, 16
; SI-NEXT:    s_or_b32 s40, s40, s47
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s52, 16
; SI-NEXT:    s_or_b32 s41, s41, s47
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s72, 16
; SI-NEXT:    s_or_b32 s14, s14, s47
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s51, 16
; SI-NEXT:    s_or_b32 s15, s15, s47
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s62, 16
; SI-NEXT:    s_or_b32 s12, s12, s47
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s50, 16
; SI-NEXT:    s_or_b32 s13, s13, s47
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s58, 16
; SI-NEXT:    s_or_b32 s10, s10, s47
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s49, 16
; SI-NEXT:    s_or_b32 s11, s11, s47
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s60, 16
; SI-NEXT:    s_or_b32 s8, s8, s47
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s48, 16
; SI-NEXT:    s_or_b32 s9, s9, s47
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s56, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_or_b32 s6, s6, s47
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s39, 16
; SI-NEXT:    s_or_b32 s4, s4, s46
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s38, 16
; SI-NEXT:    s_or_b32 s7, s7, s47
; SI-NEXT:    s_or_b32 s5, s5, s46
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB21_4:
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    s_branch .LBB21_2
;
; VI-LABEL: bitcast_v32i32_to_v64f16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB21_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB21_3
; VI-NEXT:  .LBB21_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s47, s47, 3
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:  .LBB21_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB21_4:
; VI-NEXT:    s_branch .LBB21_2
;
; GFX9-LABEL: bitcast_v32i32_to_v64f16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB21_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB21_3
; GFX9-NEXT:  .LBB21_2: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s47, s47, 3
; GFX9-NEXT:    s_add_i32 s29, s29, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s27, s27, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s25, s25, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s23, s23, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s21, s21, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s19, s19, 3
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    s_add_i32 s17, s17, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    s_add_i32 s9, s9, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    s_add_i32 s13, s13, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s15, s15, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s41, s41, 3
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    s_add_i32 s45, s45, 3
; GFX9-NEXT:  .LBB21_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB21_4:
; GFX9-NEXT:    s_branch .LBB21_2
;
; GFX11-LABEL: bitcast_v32i32_to_v64f16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB21_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB21_3
; GFX11-NEXT:  .LBB21_2: ; %cmp.true
; GFX11-NEXT:    s_add_i32 s27, s27, 3
; GFX11-NEXT:    s_add_i32 s26, s26, 3
; GFX11-NEXT:    s_add_i32 s25, s25, 3
; GFX11-NEXT:    s_add_i32 s24, s24, 3
; GFX11-NEXT:    s_add_i32 s23, s23, 3
; GFX11-NEXT:    s_add_i32 s22, s22, 3
; GFX11-NEXT:    s_add_i32 s21, s21, 3
; GFX11-NEXT:    s_add_i32 s20, s20, 3
; GFX11-NEXT:    s_add_i32 s19, s19, 3
; GFX11-NEXT:    s_add_i32 s18, s18, 3
; GFX11-NEXT:    s_add_i32 s17, s17, 3
; GFX11-NEXT:    s_add_i32 s16, s16, 3
; GFX11-NEXT:    s_add_i32 s3, s3, 3
; GFX11-NEXT:    s_add_i32 s2, s2, 3
; GFX11-NEXT:    s_add_i32 s1, s1, 3
; GFX11-NEXT:    s_add_i32 s0, s0, 3
; GFX11-NEXT:    s_add_i32 s4, s4, 3
; GFX11-NEXT:    s_add_i32 s5, s5, 3
; GFX11-NEXT:    s_add_i32 s6, s6, 3
; GFX11-NEXT:    s_add_i32 s7, s7, 3
; GFX11-NEXT:    s_add_i32 s8, s8, 3
; GFX11-NEXT:    s_add_i32 s9, s9, 3
; GFX11-NEXT:    s_add_i32 s10, s10, 3
; GFX11-NEXT:    s_add_i32 s11, s11, 3
; GFX11-NEXT:    s_add_i32 s12, s12, 3
; GFX11-NEXT:    s_add_i32 s13, s13, 3
; GFX11-NEXT:    s_add_i32 s14, s14, 3
; GFX11-NEXT:    s_add_i32 s15, s15, 3
; GFX11-NEXT:    s_add_i32 s40, s40, 3
; GFX11-NEXT:    s_add_i32 s41, s41, 3
; GFX11-NEXT:    s_add_i32 s29, s29, 3
; GFX11-NEXT:    s_add_i32 s28, s28, 3
; GFX11-NEXT:  .LBB21_3: ; %end
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB21_4:
; GFX11-NEXT:    s_branch .LBB21_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define <32 x i32> @bitcast_v64f16_to_v32i32(<64 x half> %a, i32 %b) {
; SI-LABEL: bitcast_v64f16_to_v32i32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v30
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v29
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v27
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v25
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v24
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v23
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v22
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v21
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v20
; SI-NEXT:    v_mov_b32_e32 v52, v18
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
; SI-NEXT:    v_mov_b32_e32 v53, v17
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v52
; SI-NEXT:    v_mov_b32_e32 v54, v16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v53
; SI-NEXT:    v_mov_b32_e32 v55, v15
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v54
; SI-NEXT:    v_mov_b32_e32 v40, v14
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v55
; SI-NEXT:    v_mov_b32_e32 v41, v13
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v40
; SI-NEXT:    v_mov_b32_e32 v42, v12
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v28
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v42
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v61
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v62
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB22_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v63
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v61
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v32
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v36
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v49
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v33
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v48
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v35
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v39
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v38
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v37
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v42
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v41
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v40
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v55
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v53
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v52
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB22_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB22_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, v63
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v50
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v62
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v61
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_or_b32_e32 v1, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v59
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v58
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v57
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v56
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v49
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v47
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v46
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v33
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v45
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v44
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v48
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v43
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v42
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v35
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v41
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v40
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v55
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v54
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v34
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v53
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v19, v52
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v38
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v37
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v51
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:  .LBB22_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v32i32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB22_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x200
; VI-NEXT:    v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v15, 0x200, v15
; VI-NEXT:    v_or_b32_e32 v15, v15, v33
; VI-NEXT:    v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v14, 0x200, v14
; VI-NEXT:    v_or_b32_e32 v14, v14, v33
; VI-NEXT:    v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v13, 0x200, v13
; VI-NEXT:    v_or_b32_e32 v13, v13, v33
; VI-NEXT:    v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v12, 0x200, v12
; VI-NEXT:    v_or_b32_e32 v12, v12, v33
; VI-NEXT:    v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v11, 0x200, v11
; VI-NEXT:    v_or_b32_e32 v11, v11, v33
; VI-NEXT:    v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v10, 0x200, v10
; VI-NEXT:    v_or_b32_e32 v10, v10, v33
; VI-NEXT:    v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v9, 0x200, v9
; VI-NEXT:    v_or_b32_e32 v9, v9, v33
; VI-NEXT:    v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v8, 0x200, v8
; VI-NEXT:    v_or_b32_e32 v8, v8, v33
; VI-NEXT:    v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v7, 0x200, v7
; VI-NEXT:    v_or_b32_e32 v7, v7, v33
; VI-NEXT:    v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v6, 0x200, v6
; VI-NEXT:    v_or_b32_e32 v6, v6, v33
; VI-NEXT:    v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v5, 0x200, v5
; VI-NEXT:    v_or_b32_e32 v5, v5, v33
; VI-NEXT:    v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v4, 0x200, v4
; VI-NEXT:    v_or_b32_e32 v4, v4, v33
; VI-NEXT:    v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, 0x200, v3
; VI-NEXT:    v_or_b32_e32 v3, v3, v33
; VI-NEXT:    v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v2, 0x200, v2
; VI-NEXT:    v_or_b32_e32 v2, v2, v33
; VI-NEXT:    v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, 0x200, v1
; VI-NEXT:    v_or_b32_e32 v1, v1, v33
; VI-NEXT:    v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, 0x200, v0
; VI-NEXT:    v_or_b32_e32 v0, v0, v33
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v31, 0x200, v31
; VI-NEXT:    v_or_b32_e32 v31, v31, v33
; VI-NEXT:    v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v30, 0x200, v30
; VI-NEXT:    v_or_b32_e32 v30, v30, v33
; VI-NEXT:    v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v29, 0x200, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v28, 0x200, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v33
; VI-NEXT:    v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v27, 0x200, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v33
; VI-NEXT:    v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v26, 0x200, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v33
; VI-NEXT:    v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v25, 0x200, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v33
; VI-NEXT:    v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v24, 0x200, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v33
; VI-NEXT:    v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v23, 0x200, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v33
; VI-NEXT:    v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v22, 0x200, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v33
; VI-NEXT:    v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v21, 0x200, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v33
; VI-NEXT:    v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v20, 0x200, v20
; VI-NEXT:    v_or_b32_e32 v20, v20, v33
; VI-NEXT:    v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v19, 0x200, v19
; VI-NEXT:    v_or_b32_e32 v19, v19, v33
; VI-NEXT:    v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, 0x200, v18
; VI-NEXT:    v_or_b32_e32 v18, v18, v33
; VI-NEXT:    v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, 0x200, v17
; VI-NEXT:    v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, 0x200, v16
; VI-NEXT:    v_or_b32_e32 v17, v17, v33
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB22_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v32i32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB22_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_movk_i32 s6, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB22_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v32i32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB22_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-NEXT:  .LBB22_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define inreg <32 x i32> @bitcast_v64f16_to_v32i32_scalar(<64 x half> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64f16_to_v32i32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v34, s30, 0
; SI-NEXT:    v_writelane_b32 v34, s31, 1
; SI-NEXT:    v_writelane_b32 v34, s34, 2
; SI-NEXT:    v_writelane_b32 v34, s35, 3
; SI-NEXT:    v_writelane_b32 v34, s36, 4
; SI-NEXT:    v_writelane_b32 v34, s37, 5
; SI-NEXT:    v_writelane_b32 v34, s38, 6
; SI-NEXT:    v_writelane_b32 v34, s39, 7
; SI-NEXT:    v_writelane_b32 v34, s48, 8
; SI-NEXT:    v_writelane_b32 v34, s49, 9
; SI-NEXT:    v_writelane_b32 v34, s50, 10
; SI-NEXT:    v_writelane_b32 v34, s51, 11
; SI-NEXT:    v_writelane_b32 v34, s52, 12
; SI-NEXT:    v_writelane_b32 v34, s53, 13
; SI-NEXT:    v_writelane_b32 v34, s54, 14
; SI-NEXT:    v_writelane_b32 v34, s55, 15
; SI-NEXT:    v_writelane_b32 v34, s64, 16
; SI-NEXT:    v_writelane_b32 v34, s65, 17
; SI-NEXT:    v_writelane_b32 v34, s66, 18
; SI-NEXT:    v_writelane_b32 v34, s67, 19
; SI-NEXT:    v_writelane_b32 v34, s68, 20
; SI-NEXT:    v_writelane_b32 v34, s69, 21
; SI-NEXT:    v_writelane_b32 v34, s70, 22
; SI-NEXT:    v_writelane_b32 v34, s71, 23
; SI-NEXT:    v_writelane_b32 v34, s80, 24
; SI-NEXT:    v_writelane_b32 v34, s81, 25
; SI-NEXT:    v_writelane_b32 v34, s82, 26
; SI-NEXT:    v_writelane_b32 v34, s83, 27
; SI-NEXT:    v_writelane_b32 v34, s84, 28
; SI-NEXT:    v_writelane_b32 v34, s85, 29
; SI-NEXT:    v_writelane_b32 v34, s86, 30
; SI-NEXT:    v_writelane_b32 v34, s87, 31
; SI-NEXT:    v_writelane_b32 v34, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_writelane_b32 v34, s97, 33
; SI-NEXT:    s_lshr_b32 vcc_lo, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s8, v16
; SI-NEXT:    ; implicit-def: $vgpr35 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v34, s98, 34
; SI-NEXT:    s_lshr_b32 vcc_hi, s8, 16
; SI-NEXT:    v_readfirstlane_b32 s10, v15
; SI-NEXT:    v_readfirstlane_b32 s12, v14
; SI-NEXT:    v_readfirstlane_b32 s14, v13
; SI-NEXT:    v_readfirstlane_b32 s72, v12
; SI-NEXT:    v_readfirstlane_b32 s74, v11
; SI-NEXT:    v_readfirstlane_b32 s76, v10
; SI-NEXT:    v_readfirstlane_b32 s78, v9
; SI-NEXT:    v_readfirstlane_b32 s88, v8
; SI-NEXT:    v_readfirstlane_b32 s90, v7
; SI-NEXT:    v_readfirstlane_b32 s92, v6
; SI-NEXT:    v_readfirstlane_b32 s94, v5
; SI-NEXT:    v_readfirstlane_b32 s30, v4
; SI-NEXT:    v_readfirstlane_b32 s34, v3
; SI-NEXT:    v_readfirstlane_b32 s68, v2
; SI-NEXT:    v_readfirstlane_b32 s80, v1
; SI-NEXT:    v_readfirstlane_b32 s83, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v35, vcc_lo, 0
; SI-NEXT:    v_writelane_b32 v34, s99, 35
; SI-NEXT:    s_lshr_b32 s69, s29, 16
; SI-NEXT:    s_lshr_b32 s71, s28, 16
; SI-NEXT:    s_lshr_b32 s82, s27, 16
; SI-NEXT:    s_lshr_b32 s85, s26, 16
; SI-NEXT:    s_lshr_b32 s86, s25, 16
; SI-NEXT:    s_lshr_b32 s87, s24, 16
; SI-NEXT:    s_lshr_b32 s96, s23, 16
; SI-NEXT:    s_lshr_b32 s97, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s7, s19, 16
; SI-NEXT:    s_lshr_b32 s9, s18, 16
; SI-NEXT:    s_lshr_b32 s11, s17, 16
; SI-NEXT:    s_lshr_b32 s13, s16, 16
; SI-NEXT:    s_lshr_b32 s65, s10, 16
; SI-NEXT:    s_lshr_b32 s64, s12, 16
; SI-NEXT:    s_lshr_b32 s15, s14, 16
; SI-NEXT:    s_lshr_b32 s73, s72, 16
; SI-NEXT:    s_lshr_b32 s75, s74, 16
; SI-NEXT:    s_lshr_b32 s77, s76, 16
; SI-NEXT:    s_lshr_b32 s79, s78, 16
; SI-NEXT:    s_lshr_b32 s89, s88, 16
; SI-NEXT:    s_lshr_b32 s91, s90, 16
; SI-NEXT:    s_lshr_b32 s93, s92, 16
; SI-NEXT:    s_lshr_b32 s95, s94, 16
; SI-NEXT:    s_lshr_b32 s31, s30, 16
; SI-NEXT:    s_lshr_b32 s35, s34, 16
; SI-NEXT:    s_lshr_b32 s70, s68, 16
; SI-NEXT:    s_lshr_b32 s81, s80, 16
; SI-NEXT:    s_lshr_b32 s84, s83, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v35, vcc_hi, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v35, s65, 2
; SI-NEXT:    v_writelane_b32 v35, s64, 3
; SI-NEXT:    s_cbranch_scc0 .LBB23_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s13, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s9, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s7, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s11, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s97, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s86, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s34, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s30, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s31, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s92, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s90, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s91, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s88, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s89, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s78, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s79, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s76, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s77, 16
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s74, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s75, 16
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s72, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s73, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s15, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s64, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s65, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB23_4
; SI-NEXT:  .LBB23_2: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s13
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s16
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s11
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s17
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s18
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s9
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v1, v3, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s7
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s19
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s99
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s20
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s98
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s21
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s97
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v7, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s22
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s96
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s23
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s87
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s24
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_or_b32_e32 v7, v9, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s86
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v10, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s25
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s85
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s26
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s82
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s27
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_or_b32_e32 v10, v12, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s71
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v13, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s28
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s69
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s29
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s84
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s83
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_or_b32_e32 v13, v15, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s81
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s80
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s70
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s68
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s35
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s34
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_or_b32_e32 v16, v18, v16
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s31
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s30
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s95
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s94
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s93
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s92
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s91
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s90
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s89
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s88
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s79
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s78
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s77
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s76
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s75
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s74
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s73
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s72
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s15
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_readlane_b32 s4, v35, 3
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s14
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_readlane_b32 s4, v35, 2
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s12
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s10
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_readlane_b32 s4, v35, 1
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s4
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_readlane_b32 s4, v35, 0
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s8
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v33, s6
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:    s_branch .LBB23_5
; SI-NEXT:  .LBB23_3:
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB23_2
; SI-NEXT:  .LBB23_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB23_5: ; %end
; SI-NEXT:    v_readlane_b32 s99, v34, 35
; SI-NEXT:    v_readlane_b32 s98, v34, 34
; SI-NEXT:    v_readlane_b32 s97, v34, 33
; SI-NEXT:    v_readlane_b32 s96, v34, 32
; SI-NEXT:    v_readlane_b32 s87, v34, 31
; SI-NEXT:    v_readlane_b32 s86, v34, 30
; SI-NEXT:    v_readlane_b32 s85, v34, 29
; SI-NEXT:    v_readlane_b32 s84, v34, 28
; SI-NEXT:    v_readlane_b32 s83, v34, 27
; SI-NEXT:    v_readlane_b32 s82, v34, 26
; SI-NEXT:    v_readlane_b32 s81, v34, 25
; SI-NEXT:    v_readlane_b32 s80, v34, 24
; SI-NEXT:    v_readlane_b32 s71, v34, 23
; SI-NEXT:    v_readlane_b32 s70, v34, 22
; SI-NEXT:    v_readlane_b32 s69, v34, 21
; SI-NEXT:    v_readlane_b32 s68, v34, 20
; SI-NEXT:    v_readlane_b32 s67, v34, 19
; SI-NEXT:    v_readlane_b32 s66, v34, 18
; SI-NEXT:    v_readlane_b32 s65, v34, 17
; SI-NEXT:    v_readlane_b32 s64, v34, 16
; SI-NEXT:    v_readlane_b32 s55, v34, 15
; SI-NEXT:    v_readlane_b32 s54, v34, 14
; SI-NEXT:    v_readlane_b32 s53, v34, 13
; SI-NEXT:    v_readlane_b32 s52, v34, 12
; SI-NEXT:    v_readlane_b32 s51, v34, 11
; SI-NEXT:    v_readlane_b32 s50, v34, 10
; SI-NEXT:    v_readlane_b32 s49, v34, 9
; SI-NEXT:    v_readlane_b32 s48, v34, 8
; SI-NEXT:    v_readlane_b32 s39, v34, 7
; SI-NEXT:    v_readlane_b32 s38, v34, 6
; SI-NEXT:    v_readlane_b32 s37, v34, 5
; SI-NEXT:    v_readlane_b32 s36, v34, 4
; SI-NEXT:    v_readlane_b32 s35, v34, 3
; SI-NEXT:    v_readlane_b32 s34, v34, 2
; SI-NEXT:    v_readlane_b32 s31, v34, 1
; SI-NEXT:    v_readlane_b32 s30, v34, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v32i32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v33, s36, 0
; VI-NEXT:    v_writelane_b32 v33, s37, 1
; VI-NEXT:    v_writelane_b32 v33, s38, 2
; VI-NEXT:    v_writelane_b32 v33, s39, 3
; VI-NEXT:    v_writelane_b32 v33, s48, 4
; VI-NEXT:    v_writelane_b32 v33, s49, 5
; VI-NEXT:    v_writelane_b32 v33, s50, 6
; VI-NEXT:    v_writelane_b32 v33, s51, 7
; VI-NEXT:    v_writelane_b32 v33, s52, 8
; VI-NEXT:    v_writelane_b32 v33, s53, 9
; VI-NEXT:    v_writelane_b32 v33, s54, 10
; VI-NEXT:    v_writelane_b32 v33, s55, 11
; VI-NEXT:    v_writelane_b32 v33, s64, 12
; VI-NEXT:    v_writelane_b32 v33, s65, 13
; VI-NEXT:    v_writelane_b32 v33, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v33, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB23_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB23_4
; VI-NEXT:  .LBB23_2: ; %cmp.true
; VI-NEXT:    s_lshr_b32 s4, s51, 16
; VI-NEXT:    v_mov_b32_e32 v16, 0x200
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    s_lshr_b32 s4, s50, 16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s51, v16
; VI-NEXT:    v_mov_b32_e32 v2, s4
; VI-NEXT:    s_lshr_b32 s4, s49, 16
; VI-NEXT:    v_or_b32_e32 v15, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s49, v16
; VI-NEXT:    s_lshr_b32 s4, s48, 16
; VI-NEXT:    v_or_b32_e32 v13, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s48, v16
; VI-NEXT:    s_lshr_b32 s4, s47, 16
; VI-NEXT:    v_or_b32_e32 v12, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s47, v16
; VI-NEXT:    s_lshr_b32 s4, s46, 16
; VI-NEXT:    v_or_b32_e32 v11, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s46, v16
; VI-NEXT:    s_lshr_b32 s4, s45, 16
; VI-NEXT:    v_or_b32_e32 v10, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s45, v16
; VI-NEXT:    s_lshr_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v9, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s44, v16
; VI-NEXT:    s_lshr_b32 s4, s43, 16
; VI-NEXT:    v_or_b32_e32 v8, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s43, v16
; VI-NEXT:    s_lshr_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v7, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s42, v16
; VI-NEXT:    s_lshr_b32 s4, s41, 16
; VI-NEXT:    v_or_b32_e32 v6, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s41, v16
; VI-NEXT:    s_lshr_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v5, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s40, v16
; VI-NEXT:    s_lshr_b32 s4, s39, 16
; VI-NEXT:    v_or_b32_e32 v4, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v2, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, s50, v16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s39, v16
; VI-NEXT:    s_lshr_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v14, v3, v2
; VI-NEXT:    v_or_b32_e32 v3, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s38, v16
; VI-NEXT:    s_lshr_b32 s4, s37, 16
; VI-NEXT:    v_or_b32_e32 v2, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s37, v16
; VI-NEXT:    s_lshr_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v1, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, s36, v16
; VI-NEXT:    s_lshr_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v0, v17, v0
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s67, v16
; VI-NEXT:    s_lshr_b32 s4, s66, 16
; VI-NEXT:    v_or_b32_e32 v31, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s66, v16
; VI-NEXT:    s_lshr_b32 s4, s65, 16
; VI-NEXT:    v_or_b32_e32 v30, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s65, v16
; VI-NEXT:    s_lshr_b32 s4, s64, 16
; VI-NEXT:    v_or_b32_e32 v29, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s64, v16
; VI-NEXT:    s_lshr_b32 s4, s63, 16
; VI-NEXT:    v_or_b32_e32 v28, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s63, v16
; VI-NEXT:    s_lshr_b32 s4, s62, 16
; VI-NEXT:    v_or_b32_e32 v27, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s62, v16
; VI-NEXT:    s_lshr_b32 s4, s61, 16
; VI-NEXT:    v_or_b32_e32 v26, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s61, v16
; VI-NEXT:    s_lshr_b32 s4, s60, 16
; VI-NEXT:    v_or_b32_e32 v25, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s60, v16
; VI-NEXT:    s_lshr_b32 s4, s59, 16
; VI-NEXT:    v_or_b32_e32 v24, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s59, v16
; VI-NEXT:    s_lshr_b32 s4, s58, 16
; VI-NEXT:    v_or_b32_e32 v23, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s58, v16
; VI-NEXT:    s_lshr_b32 s4, s57, 16
; VI-NEXT:    v_or_b32_e32 v22, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s57, v16
; VI-NEXT:    s_lshr_b32 s4, s56, 16
; VI-NEXT:    v_or_b32_e32 v21, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s56, v16
; VI-NEXT:    s_lshr_b32 s4, s55, 16
; VI-NEXT:    v_or_b32_e32 v20, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s55, v16
; VI-NEXT:    s_lshr_b32 s4, s54, 16
; VI-NEXT:    v_or_b32_e32 v19, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s54, v16
; VI-NEXT:    s_lshr_b32 s4, s53, 16
; VI-NEXT:    v_or_b32_e32 v18, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v32, s53, v16
; VI-NEXT:    s_lshr_b32 s4, s52, 16
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_mov_b32_e32 v32, s4
; VI-NEXT:    v_add_f16_sdwa v32, v32, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, s52, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:    s_branch .LBB23_5
; VI-NEXT:  .LBB23_3:
; VI-NEXT:    s_branch .LBB23_2
; VI-NEXT:  .LBB23_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB23_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v33, 15
; VI-NEXT:    v_readlane_b32 s66, v33, 14
; VI-NEXT:    v_readlane_b32 s65, v33, 13
; VI-NEXT:    v_readlane_b32 s64, v33, 12
; VI-NEXT:    v_readlane_b32 s55, v33, 11
; VI-NEXT:    v_readlane_b32 s54, v33, 10
; VI-NEXT:    v_readlane_b32 s53, v33, 9
; VI-NEXT:    v_readlane_b32 s52, v33, 8
; VI-NEXT:    v_readlane_b32 s51, v33, 7
; VI-NEXT:    v_readlane_b32 s50, v33, 6
; VI-NEXT:    v_readlane_b32 s49, v33, 5
; VI-NEXT:    v_readlane_b32 s48, v33, 4
; VI-NEXT:    v_readlane_b32 s39, v33, 3
; VI-NEXT:    v_readlane_b32 s38, v33, 2
; VI-NEXT:    v_readlane_b32 s37, v33, 1
; VI-NEXT:    v_readlane_b32 s36, v33, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v32i32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB23_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB23_4
; GFX9-NEXT:  .LBB23_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, s51, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, s50, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, s49, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, s48, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, s47, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, s46, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, s45, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, s44, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, s43, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, s42, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, s41, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, s40, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, s39, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, s38, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, s37, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, s36, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v31, s67, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, s66, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, s65, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, s64, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, s63, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, s62, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, s61, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, s60, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, s59, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, s58, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, s57, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, s56, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, s55, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, s54, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, s53, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, s52, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB23_5
; GFX9-NEXT:  .LBB23_3:
; GFX9-NEXT:    s_branch .LBB23_2
; GFX9-NEXT:  .LBB23_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB23_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v32i32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB23_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB23_4
; GFX11-NEXT:  .LBB23_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s51 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s50 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s49 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s48 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s47 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s46 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s45 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s44 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s43 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s42 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s41 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s40 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s39 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s38 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s37 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s36 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, s67 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, s66 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s65 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s64 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, s63 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, s62 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, s61 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s60 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s59 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, s58 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s57 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s56 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s55 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s54 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s53 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s52 op_sel_hi:[0,1]
; GFX11-NEXT:    s_branch .LBB23_5
; GFX11-NEXT:  .LBB23_3:
; GFX11-NEXT:    s_branch .LBB23_2
; GFX11-NEXT:  .LBB23_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB23_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define <64 x i16> @bitcast_v32i32_to_v64i16(<32 x i32> %a, i32 %b) {
; SI-LABEL: bitcast_v32i32_to_v64i16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB24_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB24_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB24_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB24_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
; SI-NEXT:    v_or_b32_e32 v0, v0, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v44
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v42
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v61
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v47
; SI-NEXT:    v_or_b32_e32 v3, v3, v44
; SI-NEXT:    v_or_b32_e32 v5, v5, v42
; SI-NEXT:    v_or_b32_e32 v6, v6, v55
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v60
; SI-NEXT:    v_or_b32_e32 v8, v8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v57
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v45
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v41
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v54
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v52
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v53
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32i32_to_v64i16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB24_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:  .LBB24_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32i32_to_v64i16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB24_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v15
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v14
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v13
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_add_u32_e32 v11, 3, v11
; GFX9-NEXT:    v_add_u32_e32 v10, 3, v10
; GFX9-NEXT:    v_add_u32_e32 v9, 3, v9
; GFX9-NEXT:    v_add_u32_e32 v8, 3, v8
; GFX9-NEXT:    v_add_u32_e32 v7, 3, v7
; GFX9-NEXT:    v_add_u32_e32 v6, 3, v6
; GFX9-NEXT:    v_add_u32_e32 v5, 3, v5
; GFX9-NEXT:    v_add_u32_e32 v4, 3, v4
; GFX9-NEXT:    v_add_u32_e32 v3, 3, v3
; GFX9-NEXT:    v_add_u32_e32 v2, 3, v2
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v18
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v17
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v16
; GFX9-NEXT:  .LBB24_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32i32_to_v64i16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB24_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_nc_u32_e32 v15, 3, v15
; GFX11-NEXT:    v_add_nc_u32_e32 v14, 3, v14
; GFX11-NEXT:    v_add_nc_u32_e32 v13, 3, v13
; GFX11-NEXT:    v_add_nc_u32_e32 v12, 3, v12
; GFX11-NEXT:    v_add_nc_u32_e32 v11, 3, v11
; GFX11-NEXT:    v_add_nc_u32_e32 v10, 3, v10
; GFX11-NEXT:    v_add_nc_u32_e32 v9, 3, v9
; GFX11-NEXT:    v_add_nc_u32_e32 v8, 3, v8
; GFX11-NEXT:    v_add_nc_u32_e32 v7, 3, v7
; GFX11-NEXT:    v_add_nc_u32_e32 v6, 3, v6
; GFX11-NEXT:    v_add_nc_u32_e32 v5, 3, v5
; GFX11-NEXT:    v_add_nc_u32_e32 v4, 3, v4
; GFX11-NEXT:    v_add_nc_u32_e32 v3, 3, v3
; GFX11-NEXT:    v_add_nc_u32_e32 v2, 3, v2
; GFX11-NEXT:    v_add_nc_u32_e32 v1, 3, v1
; GFX11-NEXT:    v_add_nc_u32_e32 v0, 3, v0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_nc_u32_e32 v31, 3, v31
; GFX11-NEXT:    v_add_nc_u32_e32 v30, 3, v30
; GFX11-NEXT:    v_add_nc_u32_e32 v29, 3, v29
; GFX11-NEXT:    v_add_nc_u32_e32 v28, 3, v28
; GFX11-NEXT:    v_add_nc_u32_e32 v27, 3, v27
; GFX11-NEXT:    v_add_nc_u32_e32 v26, 3, v26
; GFX11-NEXT:    v_add_nc_u32_e32 v25, 3, v25
; GFX11-NEXT:    v_add_nc_u32_e32 v24, 3, v24
; GFX11-NEXT:    v_add_nc_u32_e32 v23, 3, v23
; GFX11-NEXT:    v_add_nc_u32_e32 v22, 3, v22
; GFX11-NEXT:    v_add_nc_u32_e32 v21, 3, v21
; GFX11-NEXT:    v_add_nc_u32_e32 v20, 3, v20
; GFX11-NEXT:    v_add_nc_u32_e32 v19, 3, v19
; GFX11-NEXT:    v_add_nc_u32_e32 v18, 3, v18
; GFX11-NEXT:    v_add_nc_u32_e32 v17, 3, v17
; GFX11-NEXT:    v_add_nc_u32_e32 v16, 3, v16
; GFX11-NEXT:  .LBB24_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define inreg <64 x i16> @bitcast_v32i32_to_v64i16_scalar(<32 x i32> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32i32_to_v64i16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_readfirstlane_b32 s44, v18
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_readfirstlane_b32 s7, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    v_readfirstlane_b32 s9, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_readfirstlane_b32 s10, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s15, v7
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s40, v4
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    v_readfirstlane_b32 s42, v2
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    s_cbranch_scc0 .LBB25_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB25_3
; SI-NEXT:  .LBB25_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s44, s44, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s42, s42, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s40, s40, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s10, s10, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s6, s6, 3
; SI-NEXT:    s_add_i32 s5, s5, 3
; SI-NEXT:    s_add_i32 s4, s4, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:  .LBB25_3: ; %end
; SI-NEXT:    s_lshl_b32 s47, s36, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s69, 16
; SI-NEXT:    s_or_b32 s17, s17, s47
; SI-NEXT:    s_lshl_b32 s47, s34, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s47
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s68, 16
; SI-NEXT:    s_or_b32 s19, s19, s47
; SI-NEXT:    s_lshl_b32 s47, s30, 16
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s47
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s67, 16
; SI-NEXT:    s_or_b32 s21, s21, s47
; SI-NEXT:    s_lshl_b32 s47, s94, 16
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s47
; SI-NEXT:    s_and_b32 s23, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s66, 16
; SI-NEXT:    s_or_b32 s23, s23, s47
; SI-NEXT:    s_lshl_b32 s47, s92, 16
; SI-NEXT:    s_and_b32 s24, s24, 0xffff
; SI-NEXT:    s_or_b32 s24, s24, s47
; SI-NEXT:    s_and_b32 s25, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s65, 16
; SI-NEXT:    s_or_b32 s25, s25, s47
; SI-NEXT:    s_and_b32 s26, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s90, 16
; SI-NEXT:    s_or_b32 s26, s26, s47
; SI-NEXT:    s_and_b32 s27, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s64, 16
; SI-NEXT:    s_or_b32 s27, s27, s47
; SI-NEXT:    s_and_b32 s28, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s88, 16
; SI-NEXT:    s_or_b32 s28, s28, s47
; SI-NEXT:    s_and_b32 s29, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s55, 16
; SI-NEXT:    s_or_b32 s29, s29, s47
; SI-NEXT:    s_and_b32 s44, s44, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s78, 16
; SI-NEXT:    s_or_b32 s44, s44, s47
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s54, 16
; SI-NEXT:    s_or_b32 s45, s45, s47
; SI-NEXT:    s_and_b32 s42, s42, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s76, 16
; SI-NEXT:    s_or_b32 s42, s42, s47
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s53, 16
; SI-NEXT:    s_or_b32 s43, s43, s47
; SI-NEXT:    s_and_b32 s40, s40, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s74, 16
; SI-NEXT:    s_or_b32 s40, s40, s47
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s52, 16
; SI-NEXT:    s_or_b32 s41, s41, s47
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s72, 16
; SI-NEXT:    s_or_b32 s14, s14, s47
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s51, 16
; SI-NEXT:    s_or_b32 s15, s15, s47
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s62, 16
; SI-NEXT:    s_or_b32 s12, s12, s47
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s50, 16
; SI-NEXT:    s_or_b32 s13, s13, s47
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s58, 16
; SI-NEXT:    s_or_b32 s10, s10, s47
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s49, 16
; SI-NEXT:    s_or_b32 s11, s11, s47
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s60, 16
; SI-NEXT:    s_or_b32 s8, s8, s47
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s48, 16
; SI-NEXT:    s_or_b32 s9, s9, s47
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s56, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_or_b32 s6, s6, s47
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s39, 16
; SI-NEXT:    s_or_b32 s4, s4, s46
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s38, 16
; SI-NEXT:    s_or_b32 s7, s7, s47
; SI-NEXT:    s_or_b32 s5, s5, s46
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB25_4:
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    s_branch .LBB25_2
;
; VI-LABEL: bitcast_v32i32_to_v64i16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB25_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB25_3
; VI-NEXT:  .LBB25_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s47, s47, 3
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:  .LBB25_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB25_4:
; VI-NEXT:    s_branch .LBB25_2
;
; GFX9-LABEL: bitcast_v32i32_to_v64i16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB25_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB25_3
; GFX9-NEXT:  .LBB25_2: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s47, s47, 3
; GFX9-NEXT:    s_add_i32 s29, s29, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s27, s27, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s25, s25, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s23, s23, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s21, s21, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s19, s19, 3
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    s_add_i32 s17, s17, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    s_add_i32 s9, s9, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    s_add_i32 s13, s13, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s15, s15, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s41, s41, 3
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    s_add_i32 s45, s45, 3
; GFX9-NEXT:  .LBB25_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB25_4:
; GFX9-NEXT:    s_branch .LBB25_2
;
; GFX11-LABEL: bitcast_v32i32_to_v64i16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB25_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB25_3
; GFX11-NEXT:  .LBB25_2: ; %cmp.true
; GFX11-NEXT:    s_add_i32 s27, s27, 3
; GFX11-NEXT:    s_add_i32 s26, s26, 3
; GFX11-NEXT:    s_add_i32 s25, s25, 3
; GFX11-NEXT:    s_add_i32 s24, s24, 3
; GFX11-NEXT:    s_add_i32 s23, s23, 3
; GFX11-NEXT:    s_add_i32 s22, s22, 3
; GFX11-NEXT:    s_add_i32 s21, s21, 3
; GFX11-NEXT:    s_add_i32 s20, s20, 3
; GFX11-NEXT:    s_add_i32 s19, s19, 3
; GFX11-NEXT:    s_add_i32 s18, s18, 3
; GFX11-NEXT:    s_add_i32 s17, s17, 3
; GFX11-NEXT:    s_add_i32 s16, s16, 3
; GFX11-NEXT:    s_add_i32 s3, s3, 3
; GFX11-NEXT:    s_add_i32 s2, s2, 3
; GFX11-NEXT:    s_add_i32 s1, s1, 3
; GFX11-NEXT:    s_add_i32 s0, s0, 3
; GFX11-NEXT:    s_add_i32 s4, s4, 3
; GFX11-NEXT:    s_add_i32 s5, s5, 3
; GFX11-NEXT:    s_add_i32 s6, s6, 3
; GFX11-NEXT:    s_add_i32 s7, s7, 3
; GFX11-NEXT:    s_add_i32 s8, s8, 3
; GFX11-NEXT:    s_add_i32 s9, s9, 3
; GFX11-NEXT:    s_add_i32 s10, s10, 3
; GFX11-NEXT:    s_add_i32 s11, s11, 3
; GFX11-NEXT:    s_add_i32 s12, s12, 3
; GFX11-NEXT:    s_add_i32 s13, s13, 3
; GFX11-NEXT:    s_add_i32 s14, s14, 3
; GFX11-NEXT:    s_add_i32 s15, s15, 3
; GFX11-NEXT:    s_add_i32 s40, s40, 3
; GFX11-NEXT:    s_add_i32 s41, s41, 3
; GFX11-NEXT:    s_add_i32 s29, s29, 3
; GFX11-NEXT:    s_add_i32 s28, s28, 3
; GFX11-NEXT:  .LBB25_3: ; %end
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB25_4:
; GFX11-NEXT:    s_branch .LBB25_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <32 x i32> %a, splat (i32 3)
  %a2 = bitcast <32 x i32> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <32 x i32> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define <32 x i32> @bitcast_v64i16_to_v32i32(<64 x i16> %a, i32 %b) {
; SI-LABEL: bitcast_v64i16_to_v32i32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v19
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v20
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v21
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v22
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v23
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v24
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v25
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v27
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v17
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v61
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v23
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v22
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v21
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v19
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v18
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v17
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v10
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v62
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v28
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v29
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v30
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v63
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB26_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_or_b32_e32 v8, v8, v32
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    v_or_b32_e32 v0, v0, v49
; SI-NEXT:    v_or_b32_e32 v1, v1, v42
; SI-NEXT:    v_or_b32_e32 v2, v2, v34
; SI-NEXT:    v_or_b32_e32 v3, v3, v41
; SI-NEXT:    v_or_b32_e32 v4, v4, v48
; SI-NEXT:    v_or_b32_e32 v5, v5, v40
; SI-NEXT:    v_or_b32_e32 v6, v6, v33
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v54
; SI-NEXT:    v_or_b32_e32 v10, v10, v39
; SI-NEXT:    v_or_b32_e32 v11, v11, v53
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v52
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v37
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v15, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v17, v50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v18, v35
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v63
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB26_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB26_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v46
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v32, v8
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v62
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v61
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v60
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v59
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v58
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v57
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v56
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v47
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v44
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    s_mov_b32 s6, 0x30000
; SI-NEXT:    v_or_b32_e32 v2, v34, v2
; SI-NEXT:    v_or_b32_e32 v3, v41, v3
; SI-NEXT:    v_or_b32_e32 v4, v48, v4
; SI-NEXT:    v_or_b32_e32 v5, v40, v5
; SI-NEXT:    v_or_b32_e32 v6, v33, v6
; SI-NEXT:    v_or_b32_e32 v7, v55, v7
; SI-NEXT:    v_or_b32_e32 v9, v54, v9
; SI-NEXT:    v_or_b32_e32 v10, v39, v10
; SI-NEXT:    v_or_b32_e32 v11, v53, v11
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v38, v12
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v52, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v37, v14
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v51, v15
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v36, v16
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v50, v17
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v35, v18
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
; SI-NEXT:  .LBB26_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v32i32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB26_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v33, 3
; VI-NEXT:    v_add_u16_e32 v32, 3, v15
; VI-NEXT:    v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v32, v15
; VI-NEXT:    v_add_u16_e32 v32, 3, v14
; VI-NEXT:    v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v32, v14
; VI-NEXT:    v_add_u16_e32 v32, 3, v13
; VI-NEXT:    v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v32, v13
; VI-NEXT:    v_add_u16_e32 v32, 3, v12
; VI-NEXT:    v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v32, v12
; VI-NEXT:    v_add_u16_e32 v32, 3, v11
; VI-NEXT:    v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v32, v11
; VI-NEXT:    v_add_u16_e32 v32, 3, v10
; VI-NEXT:    v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v32, v10
; VI-NEXT:    v_add_u16_e32 v32, 3, v9
; VI-NEXT:    v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v32, v9
; VI-NEXT:    v_add_u16_e32 v32, 3, v8
; VI-NEXT:    v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v32, v8
; VI-NEXT:    v_add_u16_e32 v32, 3, v7
; VI-NEXT:    v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v32, v7
; VI-NEXT:    v_add_u16_e32 v32, 3, v6
; VI-NEXT:    v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v32, v6
; VI-NEXT:    v_add_u16_e32 v32, 3, v5
; VI-NEXT:    v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v32, v5
; VI-NEXT:    v_add_u16_e32 v32, 3, v4
; VI-NEXT:    v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v32, v4
; VI-NEXT:    v_add_u16_e32 v32, 3, v3
; VI-NEXT:    v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v32, v3
; VI-NEXT:    v_add_u16_e32 v32, 3, v2
; VI-NEXT:    v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v32, v2
; VI-NEXT:    v_add_u16_e32 v32, 3, v1
; VI-NEXT:    v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v32, v1
; VI-NEXT:    v_add_u16_e32 v32, 3, v0
; VI-NEXT:    v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v32, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v31
; VI-NEXT:    v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:    v_add_u16_e32 v32, 3, v30
; VI-NEXT:    v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v32, v30
; VI-NEXT:    v_add_u16_e32 v32, 3, v29
; VI-NEXT:    v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v32, v29
; VI-NEXT:    v_add_u16_e32 v32, 3, v28
; VI-NEXT:    v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v32, v28
; VI-NEXT:    v_add_u16_e32 v32, 3, v27
; VI-NEXT:    v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v32, v27
; VI-NEXT:    v_add_u16_e32 v32, 3, v26
; VI-NEXT:    v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v32, v26
; VI-NEXT:    v_add_u16_e32 v32, 3, v25
; VI-NEXT:    v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v32, v25
; VI-NEXT:    v_add_u16_e32 v32, 3, v24
; VI-NEXT:    v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v32, v24
; VI-NEXT:    v_add_u16_e32 v32, 3, v23
; VI-NEXT:    v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v32, v23
; VI-NEXT:    v_add_u16_e32 v32, 3, v22
; VI-NEXT:    v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v32, v22
; VI-NEXT:    v_add_u16_e32 v32, 3, v21
; VI-NEXT:    v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v32, v21
; VI-NEXT:    v_add_u16_e32 v32, 3, v20
; VI-NEXT:    v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v32, v20
; VI-NEXT:    v_add_u16_e32 v32, 3, v19
; VI-NEXT:    v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v32, v19
; VI-NEXT:    v_add_u16_e32 v32, 3, v18
; VI-NEXT:    v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v32, v18
; VI-NEXT:    v_add_u16_e32 v32, 3, v17
; VI-NEXT:    v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB26_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64i16_to_v32i32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB26_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB26_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v32i32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB26_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:  .LBB26_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define inreg <32 x i32> @bitcast_v64i16_to_v32i32_scalar(<64 x i16> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64i16_to_v32i32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    v_writelane_b32 v32, s70, 22
; SI-NEXT:    v_writelane_b32 v32, s71, 23
; SI-NEXT:    v_writelane_b32 v32, s80, 24
; SI-NEXT:    v_writelane_b32 v32, s81, 25
; SI-NEXT:    v_writelane_b32 v32, s82, 26
; SI-NEXT:    v_writelane_b32 v32, s83, 27
; SI-NEXT:    v_writelane_b32 v32, s84, 28
; SI-NEXT:    v_writelane_b32 v32, s85, 29
; SI-NEXT:    v_writelane_b32 v32, s86, 30
; SI-NEXT:    v_writelane_b32 v32, s87, 31
; SI-NEXT:    v_writelane_b32 v32, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s9, v16
; SI-NEXT:    v_writelane_b32 v32, s97, 33
; SI-NEXT:    s_lshr_b32 s14, s9, 16
; SI-NEXT:    v_readfirstlane_b32 s13, v14
; SI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v32, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s7, v17
; SI-NEXT:    v_readfirstlane_b32 s11, v15
; SI-NEXT:    s_lshr_b32 s72, s13, 16
; SI-NEXT:    v_readfirstlane_b32 s15, v13
; SI-NEXT:    v_readfirstlane_b32 s73, v12
; SI-NEXT:    v_readfirstlane_b32 s75, v11
; SI-NEXT:    v_readfirstlane_b32 s77, v10
; SI-NEXT:    v_readfirstlane_b32 s79, v9
; SI-NEXT:    v_readfirstlane_b32 s89, v8
; SI-NEXT:    v_readfirstlane_b32 s91, v7
; SI-NEXT:    v_readfirstlane_b32 s94, v6
; SI-NEXT:    v_readfirstlane_b32 s31, v5
; SI-NEXT:    v_readfirstlane_b32 s68, v4
; SI-NEXT:    v_readfirstlane_b32 s80, v3
; SI-NEXT:    v_readfirstlane_b32 s83, v2
; SI-NEXT:    v_readfirstlane_b32 s86, v1
; SI-NEXT:    v_readfirstlane_b32 s97, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v33, s14, 0
; SI-NEXT:    v_writelane_b32 v32, s99, 35
; SI-NEXT:    s_lshr_b32 s92, s29, 16
; SI-NEXT:    s_lshr_b32 s95, s28, 16
; SI-NEXT:    s_lshr_b32 s34, s27, 16
; SI-NEXT:    s_lshr_b32 s69, s26, 16
; SI-NEXT:    s_lshr_b32 s71, s25, 16
; SI-NEXT:    s_lshr_b32 s81, s24, 16
; SI-NEXT:    s_lshr_b32 s84, s23, 16
; SI-NEXT:    s_lshr_b32 s87, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s6, s19, 16
; SI-NEXT:    s_lshr_b32 s8, s18, 16
; SI-NEXT:    s_lshr_b32 s10, s17, 16
; SI-NEXT:    s_lshr_b32 s12, s16, 16
; SI-NEXT:    s_lshr_b32 vcc_lo, s7, 16
; SI-NEXT:    s_lshr_b32 vcc_hi, s11, 16
; SI-NEXT:    s_lshr_b32 s63, s15, 16
; SI-NEXT:    s_lshr_b32 s62, s73, 16
; SI-NEXT:    s_lshr_b32 s61, s75, 16
; SI-NEXT:    s_lshr_b32 s60, s77, 16
; SI-NEXT:    s_lshr_b32 s78, s79, 16
; SI-NEXT:    s_lshr_b32 s88, s89, 16
; SI-NEXT:    s_lshr_b32 s90, s91, 16
; SI-NEXT:    s_lshr_b32 s93, s94, 16
; SI-NEXT:    s_lshr_b32 s30, s31, 16
; SI-NEXT:    s_lshr_b32 s35, s68, 16
; SI-NEXT:    s_lshr_b32 s70, s80, 16
; SI-NEXT:    s_lshr_b32 s82, s83, 16
; SI-NEXT:    s_lshr_b32 s85, s86, 16
; SI-NEXT:    s_lshr_b32 s96, s97, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v33, s72, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v33, s63, 2
; SI-NEXT:    v_writelane_b32 v33, s62, 3
; SI-NEXT:    s_cbranch_scc0 .LBB27_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s10, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s60, 16
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s61, 16
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s62, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s63, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB27_3
; SI-NEXT:  .LBB27_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_or_b32 s5, s10, s5
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s36, s4, 0x30000
; SI-NEXT:    s_add_i32 s37, s5, 0x30000
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s38, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s39, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s40, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s41, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s42, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s43, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s44, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s45, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s46, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s47, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s48, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s97, s97, 3
; SI-NEXT:    s_add_i32 s49, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s86, s86, 3
; SI-NEXT:    s_add_i32 s50, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s83, s83, 3
; SI-NEXT:    s_add_i32 s51, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s80, s80, 3
; SI-NEXT:    s_add_i32 s52, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s68, s68, 3
; SI-NEXT:    s_add_i32 s53, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s31, s31, 3
; SI-NEXT:    s_add_i32 s54, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s94, s94, 3
; SI-NEXT:    s_add_i32 s55, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s91, s91, 3
; SI-NEXT:    s_add_i32 s56, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s89, s89, 3
; SI-NEXT:    s_add_i32 s57, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s79, s79, 3
; SI-NEXT:    s_add_i32 s58, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s77, s77, 3
; SI-NEXT:    s_add_i32 s59, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s76, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s75, s75, 3
; SI-NEXT:    s_add_i32 s60, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s74, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s73, s73, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 3
; SI-NEXT:    s_add_i32 s61, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 2
; SI-NEXT:    s_add_i32 s62, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 1
; SI-NEXT:    s_add_i32 s63, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s64, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 0
; SI-NEXT:    s_add_i32 s65, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s66, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s67, s4, 0x30000
; SI-NEXT:  .LBB27_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:    v_readlane_b32 s99, v32, 35
; SI-NEXT:    v_readlane_b32 s98, v32, 34
; SI-NEXT:    v_readlane_b32 s97, v32, 33
; SI-NEXT:    v_readlane_b32 s96, v32, 32
; SI-NEXT:    v_readlane_b32 s87, v32, 31
; SI-NEXT:    v_readlane_b32 s86, v32, 30
; SI-NEXT:    v_readlane_b32 s85, v32, 29
; SI-NEXT:    v_readlane_b32 s84, v32, 28
; SI-NEXT:    v_readlane_b32 s83, v32, 27
; SI-NEXT:    v_readlane_b32 s82, v32, 26
; SI-NEXT:    v_readlane_b32 s81, v32, 25
; SI-NEXT:    v_readlane_b32 s80, v32, 24
; SI-NEXT:    v_readlane_b32 s71, v32, 23
; SI-NEXT:    v_readlane_b32 s70, v32, 22
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB27_4:
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB27_2
;
; VI-LABEL: bitcast_v64i16_to_v32i32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB27_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB27_3
; VI-NEXT:  .LBB27_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s5, s46, 3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s47, 3
; VI-NEXT:    s_add_i32 s46, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s29, 3
; VI-NEXT:    s_add_i32 s47, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s28, 3
; VI-NEXT:    s_add_i32 s29, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s28, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s27, 3
; VI-NEXT:    s_add_i32 s28, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s27, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s26, 3
; VI-NEXT:    s_add_i32 s27, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s26, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s25, 3
; VI-NEXT:    s_add_i32 s26, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s25, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s24, 3
; VI-NEXT:    s_add_i32 s25, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s24, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s23, 3
; VI-NEXT:    s_add_i32 s24, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s23, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s22, 3
; VI-NEXT:    s_add_i32 s23, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s22, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s21, 3
; VI-NEXT:    s_add_i32 s22, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s21, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s20, 3
; VI-NEXT:    s_add_i32 s21, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s20, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s19, 3
; VI-NEXT:    s_add_i32 s20, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s19, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s18, 3
; VI-NEXT:    s_add_i32 s19, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s18, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s17, 3
; VI-NEXT:    s_add_i32 s18, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s17, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s16, 3
; VI-NEXT:    s_add_i32 s17, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s16, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s6, 3
; VI-NEXT:    s_add_i32 s16, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s6, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s7, 3
; VI-NEXT:    s_add_i32 s6, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s7, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s8, 3
; VI-NEXT:    s_add_i32 s7, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s8, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s9, 3
; VI-NEXT:    s_add_i32 s8, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s9, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s10, 3
; VI-NEXT:    s_add_i32 s9, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s10, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s11, 3
; VI-NEXT:    s_add_i32 s10, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s11, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s12, 3
; VI-NEXT:    s_add_i32 s11, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s12, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s13, 3
; VI-NEXT:    s_add_i32 s12, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s13, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s14, 3
; VI-NEXT:    s_add_i32 s13, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s14, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s15, 3
; VI-NEXT:    s_add_i32 s14, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s15, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s40, 3
; VI-NEXT:    s_add_i32 s15, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s41, 3
; VI-NEXT:    s_add_i32 s40, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s42, 3
; VI-NEXT:    s_add_i32 s41, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s43, 3
; VI-NEXT:    s_add_i32 s42, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s44, 3
; VI-NEXT:    s_add_i32 s43, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s45, 3
; VI-NEXT:    s_add_i32 s44, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s45, s4, 0x30000
; VI-NEXT:  .LBB27_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB27_4:
; VI-NEXT:    s_branch .LBB27_2
;
; GFX9-LABEL: bitcast_v64i16_to_v32i32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB27_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB27_4
; GFX9-NEXT:  .LBB27_2: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB27_5
; GFX9-NEXT:  .LBB27_3:
; GFX9-NEXT:    s_branch .LBB27_2
; GFX9-NEXT:  .LBB27_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB27_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v32i32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB27_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB27_4
; GFX11-NEXT:  .LBB27_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_branch .LBB27_5
; GFX11-NEXT:  .LBB27_3:
; GFX11-NEXT:    s_branch .LBB27_2
; GFX11-NEXT:  .LBB27_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB27_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <32 x i32>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <32 x i32>
  br label %end

end:
  %phi = phi <32 x i32> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x i32> %phi
}

define <16 x i64> @bitcast_v32f32_to_v16i64(<32 x float> %a, i32 %b) {
; SI-LABEL: bitcast_v32f32_to_v16i64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB28_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; SI-NEXT:  .LBB28_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v16i64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB28_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; VI-NEXT:  .LBB28_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v16i64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB28_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30
; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29
; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28
; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27
; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26
; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25
; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24
; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23
; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22
; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21
; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20
; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19
; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18
; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17
; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16
; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15
; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14
; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13
; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12
; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11
; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10
; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9
; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8
; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7
; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6
; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5
; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4
; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3
; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2
; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0
; GFX9-NEXT:  .LBB28_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v16i64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB28_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
; GFX11-NEXT:  .LBB28_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define inreg <16 x i64> @bitcast_v32f32_to_v16i64_scalar(<32 x float> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32f32_to_v16i64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s36, 0
; SI-NEXT:    v_writelane_b32 v32, s37, 1
; SI-NEXT:    v_writelane_b32 v32, s38, 2
; SI-NEXT:    v_writelane_b32 v32, s39, 3
; SI-NEXT:    v_writelane_b32 v32, s48, 4
; SI-NEXT:    v_writelane_b32 v32, s49, 5
; SI-NEXT:    v_writelane_b32 v32, s50, 6
; SI-NEXT:    v_writelane_b32 v32, s51, 7
; SI-NEXT:    v_writelane_b32 v32, s52, 8
; SI-NEXT:    v_writelane_b32 v32, s53, 9
; SI-NEXT:    v_writelane_b32 v32, s54, 10
; SI-NEXT:    v_writelane_b32 v32, s55, 11
; SI-NEXT:    v_writelane_b32 v32, s64, 12
; SI-NEXT:    v_writelane_b32 v32, s65, 13
; SI-NEXT:    v_writelane_b32 v32, s66, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v32, s67, 15
; SI-NEXT:    s_mov_b32 s49, s29
; SI-NEXT:    s_mov_b32 s48, s28
; SI-NEXT:    s_mov_b32 s47, s27
; SI-NEXT:    s_mov_b32 s46, s26
; SI-NEXT:    s_mov_b32 s45, s25
; SI-NEXT:    s_mov_b32 s44, s24
; SI-NEXT:    s_mov_b32 s43, s23
; SI-NEXT:    s_mov_b32 s42, s22
; SI-NEXT:    s_mov_b32 s41, s21
; SI-NEXT:    s_mov_b32 s40, s20
; SI-NEXT:    s_mov_b32 s39, s19
; SI-NEXT:    s_mov_b32 s38, s18
; SI-NEXT:    s_mov_b32 s37, s17
; SI-NEXT:    s_mov_b32 s36, s16
; SI-NEXT:    v_readfirstlane_b32 s67, v17
; SI-NEXT:    v_readfirstlane_b32 s66, v16
; SI-NEXT:    v_readfirstlane_b32 s65, v15
; SI-NEXT:    v_readfirstlane_b32 s64, v14
; SI-NEXT:    v_readfirstlane_b32 s63, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v12
; SI-NEXT:    v_readfirstlane_b32 s61, v11
; SI-NEXT:    v_readfirstlane_b32 s60, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s58, v8
; SI-NEXT:    v_readfirstlane_b32 s57, v7
; SI-NEXT:    v_readfirstlane_b32 s56, v6
; SI-NEXT:    v_readfirstlane_b32 s55, v5
; SI-NEXT:    v_readfirstlane_b32 s54, v4
; SI-NEXT:    v_readfirstlane_b32 s53, v3
; SI-NEXT:    v_readfirstlane_b32 s52, v2
; SI-NEXT:    v_readfirstlane_b32 s51, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s50, v0
; SI-NEXT:    s_cbranch_scc0 .LBB29_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB29_4
; SI-NEXT:  .LBB29_2: ; %cmp.true
; SI-NEXT:    v_add_f32_e64 v31, s67, 1.0
; SI-NEXT:    v_add_f32_e64 v30, s66, 1.0
; SI-NEXT:    v_add_f32_e64 v29, s65, 1.0
; SI-NEXT:    v_add_f32_e64 v28, s64, 1.0
; SI-NEXT:    v_add_f32_e64 v27, s63, 1.0
; SI-NEXT:    v_add_f32_e64 v26, s62, 1.0
; SI-NEXT:    v_add_f32_e64 v25, s61, 1.0
; SI-NEXT:    v_add_f32_e64 v24, s60, 1.0
; SI-NEXT:    v_add_f32_e64 v23, s59, 1.0
; SI-NEXT:    v_add_f32_e64 v22, s58, 1.0
; SI-NEXT:    v_add_f32_e64 v21, s57, 1.0
; SI-NEXT:    v_add_f32_e64 v20, s56, 1.0
; SI-NEXT:    v_add_f32_e64 v19, s55, 1.0
; SI-NEXT:    v_add_f32_e64 v18, s54, 1.0
; SI-NEXT:    v_add_f32_e64 v17, s53, 1.0
; SI-NEXT:    v_add_f32_e64 v16, s52, 1.0
; SI-NEXT:    v_add_f32_e64 v15, s51, 1.0
; SI-NEXT:    v_add_f32_e64 v14, s50, 1.0
; SI-NEXT:    v_add_f32_e64 v13, s49, 1.0
; SI-NEXT:    v_add_f32_e64 v12, s48, 1.0
; SI-NEXT:    v_add_f32_e64 v11, s47, 1.0
; SI-NEXT:    v_add_f32_e64 v10, s46, 1.0
; SI-NEXT:    v_add_f32_e64 v9, s45, 1.0
; SI-NEXT:    v_add_f32_e64 v8, s44, 1.0
; SI-NEXT:    v_add_f32_e64 v7, s43, 1.0
; SI-NEXT:    v_add_f32_e64 v6, s42, 1.0
; SI-NEXT:    v_add_f32_e64 v5, s41, 1.0
; SI-NEXT:    v_add_f32_e64 v4, s40, 1.0
; SI-NEXT:    v_add_f32_e64 v3, s39, 1.0
; SI-NEXT:    v_add_f32_e64 v2, s38, 1.0
; SI-NEXT:    v_add_f32_e64 v1, s37, 1.0
; SI-NEXT:    v_add_f32_e64 v0, s36, 1.0
; SI-NEXT:    s_branch .LBB29_5
; SI-NEXT:  .LBB29_3:
; SI-NEXT:    s_branch .LBB29_2
; SI-NEXT:  .LBB29_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB29_5: ; %end
; SI-NEXT:    v_readlane_b32 s67, v32, 15
; SI-NEXT:    v_readlane_b32 s66, v32, 14
; SI-NEXT:    v_readlane_b32 s65, v32, 13
; SI-NEXT:    v_readlane_b32 s64, v32, 12
; SI-NEXT:    v_readlane_b32 s55, v32, 11
; SI-NEXT:    v_readlane_b32 s54, v32, 10
; SI-NEXT:    v_readlane_b32 s53, v32, 9
; SI-NEXT:    v_readlane_b32 s52, v32, 8
; SI-NEXT:    v_readlane_b32 s51, v32, 7
; SI-NEXT:    v_readlane_b32 s50, v32, 6
; SI-NEXT:    v_readlane_b32 s49, v32, 5
; SI-NEXT:    v_readlane_b32 s48, v32, 4
; SI-NEXT:    v_readlane_b32 s39, v32, 3
; SI-NEXT:    v_readlane_b32 s38, v32, 2
; SI-NEXT:    v_readlane_b32 s37, v32, 1
; SI-NEXT:    v_readlane_b32 s36, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v16i64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s36, 0
; VI-NEXT:    v_writelane_b32 v32, s37, 1
; VI-NEXT:    v_writelane_b32 v32, s38, 2
; VI-NEXT:    v_writelane_b32 v32, s39, 3
; VI-NEXT:    v_writelane_b32 v32, s48, 4
; VI-NEXT:    v_writelane_b32 v32, s49, 5
; VI-NEXT:    v_writelane_b32 v32, s50, 6
; VI-NEXT:    v_writelane_b32 v32, s51, 7
; VI-NEXT:    v_writelane_b32 v32, s52, 8
; VI-NEXT:    v_writelane_b32 v32, s53, 9
; VI-NEXT:    v_writelane_b32 v32, s54, 10
; VI-NEXT:    v_writelane_b32 v32, s55, 11
; VI-NEXT:    v_writelane_b32 v32, s64, 12
; VI-NEXT:    v_writelane_b32 v32, s65, 13
; VI-NEXT:    v_writelane_b32 v32, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB29_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB29_4
; VI-NEXT:  .LBB29_2: ; %cmp.true
; VI-NEXT:    v_add_f32_e64 v31, s67, 1.0
; VI-NEXT:    v_add_f32_e64 v30, s66, 1.0
; VI-NEXT:    v_add_f32_e64 v29, s65, 1.0
; VI-NEXT:    v_add_f32_e64 v28, s64, 1.0
; VI-NEXT:    v_add_f32_e64 v27, s63, 1.0
; VI-NEXT:    v_add_f32_e64 v26, s62, 1.0
; VI-NEXT:    v_add_f32_e64 v25, s61, 1.0
; VI-NEXT:    v_add_f32_e64 v24, s60, 1.0
; VI-NEXT:    v_add_f32_e64 v23, s59, 1.0
; VI-NEXT:    v_add_f32_e64 v22, s58, 1.0
; VI-NEXT:    v_add_f32_e64 v21, s57, 1.0
; VI-NEXT:    v_add_f32_e64 v20, s56, 1.0
; VI-NEXT:    v_add_f32_e64 v19, s55, 1.0
; VI-NEXT:    v_add_f32_e64 v18, s54, 1.0
; VI-NEXT:    v_add_f32_e64 v17, s53, 1.0
; VI-NEXT:    v_add_f32_e64 v16, s52, 1.0
; VI-NEXT:    v_add_f32_e64 v15, s51, 1.0
; VI-NEXT:    v_add_f32_e64 v14, s50, 1.0
; VI-NEXT:    v_add_f32_e64 v13, s49, 1.0
; VI-NEXT:    v_add_f32_e64 v12, s48, 1.0
; VI-NEXT:    v_add_f32_e64 v11, s47, 1.0
; VI-NEXT:    v_add_f32_e64 v10, s46, 1.0
; VI-NEXT:    v_add_f32_e64 v9, s45, 1.0
; VI-NEXT:    v_add_f32_e64 v8, s44, 1.0
; VI-NEXT:    v_add_f32_e64 v7, s43, 1.0
; VI-NEXT:    v_add_f32_e64 v6, s42, 1.0
; VI-NEXT:    v_add_f32_e64 v5, s41, 1.0
; VI-NEXT:    v_add_f32_e64 v4, s40, 1.0
; VI-NEXT:    v_add_f32_e64 v3, s39, 1.0
; VI-NEXT:    v_add_f32_e64 v2, s38, 1.0
; VI-NEXT:    v_add_f32_e64 v1, s37, 1.0
; VI-NEXT:    v_add_f32_e64 v0, s36, 1.0
; VI-NEXT:    s_branch .LBB29_5
; VI-NEXT:  .LBB29_3:
; VI-NEXT:    s_branch .LBB29_2
; VI-NEXT:  .LBB29_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB29_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v32, 15
; VI-NEXT:    v_readlane_b32 s66, v32, 14
; VI-NEXT:    v_readlane_b32 s65, v32, 13
; VI-NEXT:    v_readlane_b32 s64, v32, 12
; VI-NEXT:    v_readlane_b32 s55, v32, 11
; VI-NEXT:    v_readlane_b32 s54, v32, 10
; VI-NEXT:    v_readlane_b32 s53, v32, 9
; VI-NEXT:    v_readlane_b32 s52, v32, 8
; VI-NEXT:    v_readlane_b32 s51, v32, 7
; VI-NEXT:    v_readlane_b32 s50, v32, 6
; VI-NEXT:    v_readlane_b32 s49, v32, 5
; VI-NEXT:    v_readlane_b32 s48, v32, 4
; VI-NEXT:    v_readlane_b32 s39, v32, 3
; VI-NEXT:    v_readlane_b32 s38, v32, 2
; VI-NEXT:    v_readlane_b32 s37, v32, 1
; VI-NEXT:    v_readlane_b32 s36, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v16i64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB29_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB29_4
; GFX9-NEXT:  .LBB29_2: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e64 v31, s67, 1.0
; GFX9-NEXT:    v_add_f32_e64 v30, s66, 1.0
; GFX9-NEXT:    v_add_f32_e64 v29, s65, 1.0
; GFX9-NEXT:    v_add_f32_e64 v28, s64, 1.0
; GFX9-NEXT:    v_add_f32_e64 v27, s63, 1.0
; GFX9-NEXT:    v_add_f32_e64 v26, s62, 1.0
; GFX9-NEXT:    v_add_f32_e64 v25, s61, 1.0
; GFX9-NEXT:    v_add_f32_e64 v24, s60, 1.0
; GFX9-NEXT:    v_add_f32_e64 v23, s59, 1.0
; GFX9-NEXT:    v_add_f32_e64 v22, s58, 1.0
; GFX9-NEXT:    v_add_f32_e64 v21, s57, 1.0
; GFX9-NEXT:    v_add_f32_e64 v20, s56, 1.0
; GFX9-NEXT:    v_add_f32_e64 v19, s55, 1.0
; GFX9-NEXT:    v_add_f32_e64 v18, s54, 1.0
; GFX9-NEXT:    v_add_f32_e64 v17, s53, 1.0
; GFX9-NEXT:    v_add_f32_e64 v16, s52, 1.0
; GFX9-NEXT:    v_add_f32_e64 v15, s51, 1.0
; GFX9-NEXT:    v_add_f32_e64 v14, s50, 1.0
; GFX9-NEXT:    v_add_f32_e64 v13, s49, 1.0
; GFX9-NEXT:    v_add_f32_e64 v12, s48, 1.0
; GFX9-NEXT:    v_add_f32_e64 v11, s47, 1.0
; GFX9-NEXT:    v_add_f32_e64 v10, s46, 1.0
; GFX9-NEXT:    v_add_f32_e64 v9, s45, 1.0
; GFX9-NEXT:    v_add_f32_e64 v8, s44, 1.0
; GFX9-NEXT:    v_add_f32_e64 v7, s43, 1.0
; GFX9-NEXT:    v_add_f32_e64 v6, s42, 1.0
; GFX9-NEXT:    v_add_f32_e64 v5, s41, 1.0
; GFX9-NEXT:    v_add_f32_e64 v4, s40, 1.0
; GFX9-NEXT:    v_add_f32_e64 v3, s39, 1.0
; GFX9-NEXT:    v_add_f32_e64 v2, s38, 1.0
; GFX9-NEXT:    v_add_f32_e64 v1, s37, 1.0
; GFX9-NEXT:    v_add_f32_e64 v0, s36, 1.0
; GFX9-NEXT:    s_branch .LBB29_5
; GFX9-NEXT:  .LBB29_3:
; GFX9-NEXT:    s_branch .LBB29_2
; GFX9-NEXT:  .LBB29_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB29_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v16i64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB29_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB29_4
; GFX11-NEXT:  .LBB29_2: ; %cmp.true
; GFX11-NEXT:    v_add_f32_e64 v31, s67, 1.0
; GFX11-NEXT:    v_add_f32_e64 v30, s66, 1.0
; GFX11-NEXT:    v_add_f32_e64 v29, s65, 1.0
; GFX11-NEXT:    v_add_f32_e64 v28, s64, 1.0
; GFX11-NEXT:    v_add_f32_e64 v27, s63, 1.0
; GFX11-NEXT:    v_add_f32_e64 v26, s62, 1.0
; GFX11-NEXT:    v_add_f32_e64 v25, s61, 1.0
; GFX11-NEXT:    v_add_f32_e64 v24, s60, 1.0
; GFX11-NEXT:    v_add_f32_e64 v23, s59, 1.0
; GFX11-NEXT:    v_add_f32_e64 v22, s58, 1.0
; GFX11-NEXT:    v_add_f32_e64 v21, s57, 1.0
; GFX11-NEXT:    v_add_f32_e64 v20, s56, 1.0
; GFX11-NEXT:    v_add_f32_e64 v19, s55, 1.0
; GFX11-NEXT:    v_add_f32_e64 v18, s54, 1.0
; GFX11-NEXT:    v_add_f32_e64 v17, s53, 1.0
; GFX11-NEXT:    v_add_f32_e64 v16, s52, 1.0
; GFX11-NEXT:    v_add_f32_e64 v15, s51, 1.0
; GFX11-NEXT:    v_add_f32_e64 v14, s50, 1.0
; GFX11-NEXT:    v_add_f32_e64 v13, s49, 1.0
; GFX11-NEXT:    v_add_f32_e64 v12, s48, 1.0
; GFX11-NEXT:    v_add_f32_e64 v11, s47, 1.0
; GFX11-NEXT:    v_add_f32_e64 v10, s46, 1.0
; GFX11-NEXT:    v_add_f32_e64 v9, s45, 1.0
; GFX11-NEXT:    v_add_f32_e64 v8, s44, 1.0
; GFX11-NEXT:    v_add_f32_e64 v7, s43, 1.0
; GFX11-NEXT:    v_add_f32_e64 v6, s42, 1.0
; GFX11-NEXT:    v_add_f32_e64 v5, s41, 1.0
; GFX11-NEXT:    v_add_f32_e64 v4, s40, 1.0
; GFX11-NEXT:    v_add_f32_e64 v3, s39, 1.0
; GFX11-NEXT:    v_add_f32_e64 v2, s38, 1.0
; GFX11-NEXT:    v_add_f32_e64 v1, s37, 1.0
; GFX11-NEXT:    v_add_f32_e64 v0, s36, 1.0
; GFX11-NEXT:    s_branch .LBB29_5
; GFX11-NEXT:  .LBB29_3:
; GFX11-NEXT:    s_branch .LBB29_2
; GFX11-NEXT:  .LBB29_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB29_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define <32 x float> @bitcast_v16i64_to_v32f32(<16 x i64> %a, i32 %b) {
; SI-LABEL: bitcast_v16i64_to_v32f32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB30_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; SI-NEXT:  .LBB30_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16i64_to_v32f32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB30_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; VI-NEXT:  .LBB30_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16i64_to_v32f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB30_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v28
; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v26
; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v24
; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v22
; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v20
; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v18, vcc, 3, v18
; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v16
; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v14
; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v12
; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v10
; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v8
; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v6
; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v4
; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v2
; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v0
; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX9-NEXT:  .LBB30_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16i64_to_v32f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB30_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 3
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
; GFX11-NEXT:  .LBB30_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define inreg <32 x float> @bitcast_v16i64_to_v32f32_scalar(<16 x i64> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16i64_to_v32f32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s10, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s12, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s14, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s42, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s47, v0
; SI-NEXT:    s_cbranch_scc0 .LBB31_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB31_3
; SI-NEXT:  .LBB31_2: ; %cmp.true
; SI-NEXT:    s_add_u32 s7, s7, 3
; SI-NEXT:    s_addc_u32 s6, s6, 0
; SI-NEXT:    s_add_u32 s9, s9, 3
; SI-NEXT:    s_addc_u32 s8, s8, 0
; SI-NEXT:    s_add_u32 s11, s11, 3
; SI-NEXT:    s_addc_u32 s10, s10, 0
; SI-NEXT:    s_add_u32 s13, s13, 3
; SI-NEXT:    s_addc_u32 s12, s12, 0
; SI-NEXT:    s_add_u32 s15, s15, 3
; SI-NEXT:    s_addc_u32 s14, s14, 0
; SI-NEXT:    s_add_u32 s41, s41, 3
; SI-NEXT:    s_addc_u32 s40, s40, 0
; SI-NEXT:    s_add_u32 s43, s43, 3
; SI-NEXT:    s_addc_u32 s42, s42, 0
; SI-NEXT:    s_add_u32 s45, s45, 3
; SI-NEXT:    s_addc_u32 s44, s44, 0
; SI-NEXT:    s_add_u32 s47, s47, 3
; SI-NEXT:    s_addc_u32 s46, s46, 0
; SI-NEXT:    s_add_u32 s28, s28, 3
; SI-NEXT:    s_addc_u32 s29, s29, 0
; SI-NEXT:    s_add_u32 s26, s26, 3
; SI-NEXT:    s_addc_u32 s27, s27, 0
; SI-NEXT:    s_add_u32 s24, s24, 3
; SI-NEXT:    s_addc_u32 s25, s25, 0
; SI-NEXT:    s_add_u32 s22, s22, 3
; SI-NEXT:    s_addc_u32 s23, s23, 0
; SI-NEXT:    s_add_u32 s20, s20, 3
; SI-NEXT:    s_addc_u32 s21, s21, 0
; SI-NEXT:    s_add_u32 s18, s18, 3
; SI-NEXT:    s_addc_u32 s19, s19, 0
; SI-NEXT:    s_add_u32 s16, s16, 3
; SI-NEXT:    s_addc_u32 s17, s17, 0
; SI-NEXT:  .LBB31_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s47
; SI-NEXT:    v_mov_b32_e32 v15, s46
; SI-NEXT:    v_mov_b32_e32 v16, s45
; SI-NEXT:    v_mov_b32_e32 v17, s44
; SI-NEXT:    v_mov_b32_e32 v18, s43
; SI-NEXT:    v_mov_b32_e32 v19, s42
; SI-NEXT:    v_mov_b32_e32 v20, s41
; SI-NEXT:    v_mov_b32_e32 v21, s40
; SI-NEXT:    v_mov_b32_e32 v22, s15
; SI-NEXT:    v_mov_b32_e32 v23, s14
; SI-NEXT:    v_mov_b32_e32 v24, s13
; SI-NEXT:    v_mov_b32_e32 v25, s12
; SI-NEXT:    v_mov_b32_e32 v26, s11
; SI-NEXT:    v_mov_b32_e32 v27, s10
; SI-NEXT:    v_mov_b32_e32 v28, s9
; SI-NEXT:    v_mov_b32_e32 v29, s8
; SI-NEXT:    v_mov_b32_e32 v30, s7
; SI-NEXT:    v_mov_b32_e32 v31, s6
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB31_4:
; SI-NEXT:    s_branch .LBB31_2
;
; VI-LABEL: bitcast_v16i64_to_v32f32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB31_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB31_3
; VI-NEXT:  .LBB31_2: ; %cmp.true
; VI-NEXT:    s_add_u32 s7, s7, 3
; VI-NEXT:    s_addc_u32 s6, s6, 0
; VI-NEXT:    s_add_u32 s9, s9, 3
; VI-NEXT:    s_addc_u32 s8, s8, 0
; VI-NEXT:    s_add_u32 s11, s11, 3
; VI-NEXT:    s_addc_u32 s10, s10, 0
; VI-NEXT:    s_add_u32 s13, s13, 3
; VI-NEXT:    s_addc_u32 s12, s12, 0
; VI-NEXT:    s_add_u32 s15, s15, 3
; VI-NEXT:    s_addc_u32 s14, s14, 0
; VI-NEXT:    s_add_u32 s41, s41, 3
; VI-NEXT:    s_addc_u32 s40, s40, 0
; VI-NEXT:    s_add_u32 s43, s43, 3
; VI-NEXT:    s_addc_u32 s42, s42, 0
; VI-NEXT:    s_add_u32 s45, s45, 3
; VI-NEXT:    s_addc_u32 s44, s44, 0
; VI-NEXT:    s_add_u32 s47, s47, 3
; VI-NEXT:    s_addc_u32 s46, s46, 0
; VI-NEXT:    s_add_u32 s28, s28, 3
; VI-NEXT:    s_addc_u32 s29, s29, 0
; VI-NEXT:    s_add_u32 s26, s26, 3
; VI-NEXT:    s_addc_u32 s27, s27, 0
; VI-NEXT:    s_add_u32 s24, s24, 3
; VI-NEXT:    s_addc_u32 s25, s25, 0
; VI-NEXT:    s_add_u32 s22, s22, 3
; VI-NEXT:    s_addc_u32 s23, s23, 0
; VI-NEXT:    s_add_u32 s20, s20, 3
; VI-NEXT:    s_addc_u32 s21, s21, 0
; VI-NEXT:    s_add_u32 s18, s18, 3
; VI-NEXT:    s_addc_u32 s19, s19, 0
; VI-NEXT:    s_add_u32 s16, s16, 3
; VI-NEXT:    s_addc_u32 s17, s17, 0
; VI-NEXT:  .LBB31_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB31_4:
; VI-NEXT:    s_branch .LBB31_2
;
; GFX9-LABEL: bitcast_v16i64_to_v32f32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB31_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB31_3
; GFX9-NEXT:  .LBB31_2: ; %cmp.true
; GFX9-NEXT:    s_add_u32 s7, s7, 3
; GFX9-NEXT:    s_addc_u32 s6, s6, 0
; GFX9-NEXT:    s_add_u32 s9, s9, 3
; GFX9-NEXT:    s_addc_u32 s8, s8, 0
; GFX9-NEXT:    s_add_u32 s11, s11, 3
; GFX9-NEXT:    s_addc_u32 s10, s10, 0
; GFX9-NEXT:    s_add_u32 s13, s13, 3
; GFX9-NEXT:    s_addc_u32 s12, s12, 0
; GFX9-NEXT:    s_add_u32 s15, s15, 3
; GFX9-NEXT:    s_addc_u32 s14, s14, 0
; GFX9-NEXT:    s_add_u32 s41, s41, 3
; GFX9-NEXT:    s_addc_u32 s40, s40, 0
; GFX9-NEXT:    s_add_u32 s43, s43, 3
; GFX9-NEXT:    s_addc_u32 s42, s42, 0
; GFX9-NEXT:    s_add_u32 s45, s45, 3
; GFX9-NEXT:    s_addc_u32 s44, s44, 0
; GFX9-NEXT:    s_add_u32 s47, s47, 3
; GFX9-NEXT:    s_addc_u32 s46, s46, 0
; GFX9-NEXT:    s_add_u32 s28, s28, 3
; GFX9-NEXT:    s_addc_u32 s29, s29, 0
; GFX9-NEXT:    s_add_u32 s26, s26, 3
; GFX9-NEXT:    s_addc_u32 s27, s27, 0
; GFX9-NEXT:    s_add_u32 s24, s24, 3
; GFX9-NEXT:    s_addc_u32 s25, s25, 0
; GFX9-NEXT:    s_add_u32 s22, s22, 3
; GFX9-NEXT:    s_addc_u32 s23, s23, 0
; GFX9-NEXT:    s_add_u32 s20, s20, 3
; GFX9-NEXT:    s_addc_u32 s21, s21, 0
; GFX9-NEXT:    s_add_u32 s18, s18, 3
; GFX9-NEXT:    s_addc_u32 s19, s19, 0
; GFX9-NEXT:    s_add_u32 s16, s16, 3
; GFX9-NEXT:    s_addc_u32 s17, s17, 0
; GFX9-NEXT:  .LBB31_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB31_4:
; GFX9-NEXT:    s_branch .LBB31_2
;
; GFX11-LABEL: bitcast_v16i64_to_v32f32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB31_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB31_3
; GFX11-NEXT:  .LBB31_2: ; %cmp.true
; GFX11-NEXT:    s_add_u32 s5, s5, 3
; GFX11-NEXT:    s_addc_u32 s4, s4, 0
; GFX11-NEXT:    s_add_u32 s7, s7, 3
; GFX11-NEXT:    s_addc_u32 s6, s6, 0
; GFX11-NEXT:    s_add_u32 s9, s9, 3
; GFX11-NEXT:    s_addc_u32 s8, s8, 0
; GFX11-NEXT:    s_add_u32 s11, s11, 3
; GFX11-NEXT:    s_addc_u32 s10, s10, 0
; GFX11-NEXT:    s_add_u32 s13, s13, 3
; GFX11-NEXT:    s_addc_u32 s12, s12, 0
; GFX11-NEXT:    s_add_u32 s15, s15, 3
; GFX11-NEXT:    s_addc_u32 s14, s14, 0
; GFX11-NEXT:    s_add_u32 s41, s41, 3
; GFX11-NEXT:    s_addc_u32 s40, s40, 0
; GFX11-NEXT:    s_add_u32 s28, s28, 3
; GFX11-NEXT:    s_addc_u32 s29, s29, 0
; GFX11-NEXT:    s_add_u32 s26, s26, 3
; GFX11-NEXT:    s_addc_u32 s27, s27, 0
; GFX11-NEXT:    s_add_u32 s24, s24, 3
; GFX11-NEXT:    s_addc_u32 s25, s25, 0
; GFX11-NEXT:    s_add_u32 s22, s22, 3
; GFX11-NEXT:    s_addc_u32 s23, s23, 0
; GFX11-NEXT:    s_add_u32 s20, s20, 3
; GFX11-NEXT:    s_addc_u32 s21, s21, 0
; GFX11-NEXT:    s_add_u32 s18, s18, 3
; GFX11-NEXT:    s_addc_u32 s19, s19, 0
; GFX11-NEXT:    s_add_u32 s16, s16, 3
; GFX11-NEXT:    s_addc_u32 s17, s17, 0
; GFX11-NEXT:    s_add_u32 s2, s2, 3
; GFX11-NEXT:    s_addc_u32 s3, s3, 0
; GFX11-NEXT:    s_add_u32 s0, s0, 3
; GFX11-NEXT:    s_addc_u32 s1, s1, 0
; GFX11-NEXT:  .LBB31_3: ; %end
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB31_4:
; GFX11-NEXT:    s_branch .LBB31_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define <16 x double> @bitcast_v32f32_to_v16f64(<32 x float> %a, i32 %b) {
; SI-LABEL: bitcast_v32f32_to_v16f64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB32_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; SI-NEXT:  .LBB32_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v16f64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB32_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; VI-NEXT:  .LBB32_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v16f64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB32_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30
; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29
; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28
; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27
; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26
; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25
; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24
; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23
; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22
; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21
; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20
; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19
; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18
; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17
; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16
; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15
; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14
; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13
; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12
; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11
; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10
; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9
; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8
; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7
; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6
; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5
; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4
; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3
; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2
; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0
; GFX9-NEXT:  .LBB32_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v16f64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB32_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
; GFX11-NEXT:  .LBB32_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define inreg <16 x double> @bitcast_v32f32_to_v16f64_scalar(<32 x float> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32f32_to_v16f64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s36, 0
; SI-NEXT:    v_writelane_b32 v32, s37, 1
; SI-NEXT:    v_writelane_b32 v32, s38, 2
; SI-NEXT:    v_writelane_b32 v32, s39, 3
; SI-NEXT:    v_writelane_b32 v32, s48, 4
; SI-NEXT:    v_writelane_b32 v32, s49, 5
; SI-NEXT:    v_writelane_b32 v32, s50, 6
; SI-NEXT:    v_writelane_b32 v32, s51, 7
; SI-NEXT:    v_writelane_b32 v32, s52, 8
; SI-NEXT:    v_writelane_b32 v32, s53, 9
; SI-NEXT:    v_writelane_b32 v32, s54, 10
; SI-NEXT:    v_writelane_b32 v32, s55, 11
; SI-NEXT:    v_writelane_b32 v32, s64, 12
; SI-NEXT:    v_writelane_b32 v32, s65, 13
; SI-NEXT:    v_writelane_b32 v32, s66, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v32, s67, 15
; SI-NEXT:    s_mov_b32 s49, s29
; SI-NEXT:    s_mov_b32 s48, s28
; SI-NEXT:    s_mov_b32 s47, s27
; SI-NEXT:    s_mov_b32 s46, s26
; SI-NEXT:    s_mov_b32 s45, s25
; SI-NEXT:    s_mov_b32 s44, s24
; SI-NEXT:    s_mov_b32 s43, s23
; SI-NEXT:    s_mov_b32 s42, s22
; SI-NEXT:    s_mov_b32 s41, s21
; SI-NEXT:    s_mov_b32 s40, s20
; SI-NEXT:    s_mov_b32 s39, s19
; SI-NEXT:    s_mov_b32 s38, s18
; SI-NEXT:    s_mov_b32 s37, s17
; SI-NEXT:    s_mov_b32 s36, s16
; SI-NEXT:    v_readfirstlane_b32 s67, v17
; SI-NEXT:    v_readfirstlane_b32 s66, v16
; SI-NEXT:    v_readfirstlane_b32 s65, v15
; SI-NEXT:    v_readfirstlane_b32 s64, v14
; SI-NEXT:    v_readfirstlane_b32 s63, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v12
; SI-NEXT:    v_readfirstlane_b32 s61, v11
; SI-NEXT:    v_readfirstlane_b32 s60, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s58, v8
; SI-NEXT:    v_readfirstlane_b32 s57, v7
; SI-NEXT:    v_readfirstlane_b32 s56, v6
; SI-NEXT:    v_readfirstlane_b32 s55, v5
; SI-NEXT:    v_readfirstlane_b32 s54, v4
; SI-NEXT:    v_readfirstlane_b32 s53, v3
; SI-NEXT:    v_readfirstlane_b32 s52, v2
; SI-NEXT:    v_readfirstlane_b32 s51, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s50, v0
; SI-NEXT:    s_cbranch_scc0 .LBB33_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB33_4
; SI-NEXT:  .LBB33_2: ; %cmp.true
; SI-NEXT:    v_add_f32_e64 v31, s67, 1.0
; SI-NEXT:    v_add_f32_e64 v30, s66, 1.0
; SI-NEXT:    v_add_f32_e64 v29, s65, 1.0
; SI-NEXT:    v_add_f32_e64 v28, s64, 1.0
; SI-NEXT:    v_add_f32_e64 v27, s63, 1.0
; SI-NEXT:    v_add_f32_e64 v26, s62, 1.0
; SI-NEXT:    v_add_f32_e64 v25, s61, 1.0
; SI-NEXT:    v_add_f32_e64 v24, s60, 1.0
; SI-NEXT:    v_add_f32_e64 v23, s59, 1.0
; SI-NEXT:    v_add_f32_e64 v22, s58, 1.0
; SI-NEXT:    v_add_f32_e64 v21, s57, 1.0
; SI-NEXT:    v_add_f32_e64 v20, s56, 1.0
; SI-NEXT:    v_add_f32_e64 v19, s55, 1.0
; SI-NEXT:    v_add_f32_e64 v18, s54, 1.0
; SI-NEXT:    v_add_f32_e64 v17, s53, 1.0
; SI-NEXT:    v_add_f32_e64 v16, s52, 1.0
; SI-NEXT:    v_add_f32_e64 v15, s51, 1.0
; SI-NEXT:    v_add_f32_e64 v14, s50, 1.0
; SI-NEXT:    v_add_f32_e64 v13, s49, 1.0
; SI-NEXT:    v_add_f32_e64 v12, s48, 1.0
; SI-NEXT:    v_add_f32_e64 v11, s47, 1.0
; SI-NEXT:    v_add_f32_e64 v10, s46, 1.0
; SI-NEXT:    v_add_f32_e64 v9, s45, 1.0
; SI-NEXT:    v_add_f32_e64 v8, s44, 1.0
; SI-NEXT:    v_add_f32_e64 v7, s43, 1.0
; SI-NEXT:    v_add_f32_e64 v6, s42, 1.0
; SI-NEXT:    v_add_f32_e64 v5, s41, 1.0
; SI-NEXT:    v_add_f32_e64 v4, s40, 1.0
; SI-NEXT:    v_add_f32_e64 v3, s39, 1.0
; SI-NEXT:    v_add_f32_e64 v2, s38, 1.0
; SI-NEXT:    v_add_f32_e64 v1, s37, 1.0
; SI-NEXT:    v_add_f32_e64 v0, s36, 1.0
; SI-NEXT:    s_branch .LBB33_5
; SI-NEXT:  .LBB33_3:
; SI-NEXT:    s_branch .LBB33_2
; SI-NEXT:  .LBB33_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB33_5: ; %end
; SI-NEXT:    v_readlane_b32 s67, v32, 15
; SI-NEXT:    v_readlane_b32 s66, v32, 14
; SI-NEXT:    v_readlane_b32 s65, v32, 13
; SI-NEXT:    v_readlane_b32 s64, v32, 12
; SI-NEXT:    v_readlane_b32 s55, v32, 11
; SI-NEXT:    v_readlane_b32 s54, v32, 10
; SI-NEXT:    v_readlane_b32 s53, v32, 9
; SI-NEXT:    v_readlane_b32 s52, v32, 8
; SI-NEXT:    v_readlane_b32 s51, v32, 7
; SI-NEXT:    v_readlane_b32 s50, v32, 6
; SI-NEXT:    v_readlane_b32 s49, v32, 5
; SI-NEXT:    v_readlane_b32 s48, v32, 4
; SI-NEXT:    v_readlane_b32 s39, v32, 3
; SI-NEXT:    v_readlane_b32 s38, v32, 2
; SI-NEXT:    v_readlane_b32 s37, v32, 1
; SI-NEXT:    v_readlane_b32 s36, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v16f64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s36, 0
; VI-NEXT:    v_writelane_b32 v32, s37, 1
; VI-NEXT:    v_writelane_b32 v32, s38, 2
; VI-NEXT:    v_writelane_b32 v32, s39, 3
; VI-NEXT:    v_writelane_b32 v32, s48, 4
; VI-NEXT:    v_writelane_b32 v32, s49, 5
; VI-NEXT:    v_writelane_b32 v32, s50, 6
; VI-NEXT:    v_writelane_b32 v32, s51, 7
; VI-NEXT:    v_writelane_b32 v32, s52, 8
; VI-NEXT:    v_writelane_b32 v32, s53, 9
; VI-NEXT:    v_writelane_b32 v32, s54, 10
; VI-NEXT:    v_writelane_b32 v32, s55, 11
; VI-NEXT:    v_writelane_b32 v32, s64, 12
; VI-NEXT:    v_writelane_b32 v32, s65, 13
; VI-NEXT:    v_writelane_b32 v32, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB33_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB33_4
; VI-NEXT:  .LBB33_2: ; %cmp.true
; VI-NEXT:    v_add_f32_e64 v31, s67, 1.0
; VI-NEXT:    v_add_f32_e64 v30, s66, 1.0
; VI-NEXT:    v_add_f32_e64 v29, s65, 1.0
; VI-NEXT:    v_add_f32_e64 v28, s64, 1.0
; VI-NEXT:    v_add_f32_e64 v27, s63, 1.0
; VI-NEXT:    v_add_f32_e64 v26, s62, 1.0
; VI-NEXT:    v_add_f32_e64 v25, s61, 1.0
; VI-NEXT:    v_add_f32_e64 v24, s60, 1.0
; VI-NEXT:    v_add_f32_e64 v23, s59, 1.0
; VI-NEXT:    v_add_f32_e64 v22, s58, 1.0
; VI-NEXT:    v_add_f32_e64 v21, s57, 1.0
; VI-NEXT:    v_add_f32_e64 v20, s56, 1.0
; VI-NEXT:    v_add_f32_e64 v19, s55, 1.0
; VI-NEXT:    v_add_f32_e64 v18, s54, 1.0
; VI-NEXT:    v_add_f32_e64 v17, s53, 1.0
; VI-NEXT:    v_add_f32_e64 v16, s52, 1.0
; VI-NEXT:    v_add_f32_e64 v15, s51, 1.0
; VI-NEXT:    v_add_f32_e64 v14, s50, 1.0
; VI-NEXT:    v_add_f32_e64 v13, s49, 1.0
; VI-NEXT:    v_add_f32_e64 v12, s48, 1.0
; VI-NEXT:    v_add_f32_e64 v11, s47, 1.0
; VI-NEXT:    v_add_f32_e64 v10, s46, 1.0
; VI-NEXT:    v_add_f32_e64 v9, s45, 1.0
; VI-NEXT:    v_add_f32_e64 v8, s44, 1.0
; VI-NEXT:    v_add_f32_e64 v7, s43, 1.0
; VI-NEXT:    v_add_f32_e64 v6, s42, 1.0
; VI-NEXT:    v_add_f32_e64 v5, s41, 1.0
; VI-NEXT:    v_add_f32_e64 v4, s40, 1.0
; VI-NEXT:    v_add_f32_e64 v3, s39, 1.0
; VI-NEXT:    v_add_f32_e64 v2, s38, 1.0
; VI-NEXT:    v_add_f32_e64 v1, s37, 1.0
; VI-NEXT:    v_add_f32_e64 v0, s36, 1.0
; VI-NEXT:    s_branch .LBB33_5
; VI-NEXT:  .LBB33_3:
; VI-NEXT:    s_branch .LBB33_2
; VI-NEXT:  .LBB33_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB33_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v32, 15
; VI-NEXT:    v_readlane_b32 s66, v32, 14
; VI-NEXT:    v_readlane_b32 s65, v32, 13
; VI-NEXT:    v_readlane_b32 s64, v32, 12
; VI-NEXT:    v_readlane_b32 s55, v32, 11
; VI-NEXT:    v_readlane_b32 s54, v32, 10
; VI-NEXT:    v_readlane_b32 s53, v32, 9
; VI-NEXT:    v_readlane_b32 s52, v32, 8
; VI-NEXT:    v_readlane_b32 s51, v32, 7
; VI-NEXT:    v_readlane_b32 s50, v32, 6
; VI-NEXT:    v_readlane_b32 s49, v32, 5
; VI-NEXT:    v_readlane_b32 s48, v32, 4
; VI-NEXT:    v_readlane_b32 s39, v32, 3
; VI-NEXT:    v_readlane_b32 s38, v32, 2
; VI-NEXT:    v_readlane_b32 s37, v32, 1
; VI-NEXT:    v_readlane_b32 s36, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v16f64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB33_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB33_4
; GFX9-NEXT:  .LBB33_2: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e64 v31, s67, 1.0
; GFX9-NEXT:    v_add_f32_e64 v30, s66, 1.0
; GFX9-NEXT:    v_add_f32_e64 v29, s65, 1.0
; GFX9-NEXT:    v_add_f32_e64 v28, s64, 1.0
; GFX9-NEXT:    v_add_f32_e64 v27, s63, 1.0
; GFX9-NEXT:    v_add_f32_e64 v26, s62, 1.0
; GFX9-NEXT:    v_add_f32_e64 v25, s61, 1.0
; GFX9-NEXT:    v_add_f32_e64 v24, s60, 1.0
; GFX9-NEXT:    v_add_f32_e64 v23, s59, 1.0
; GFX9-NEXT:    v_add_f32_e64 v22, s58, 1.0
; GFX9-NEXT:    v_add_f32_e64 v21, s57, 1.0
; GFX9-NEXT:    v_add_f32_e64 v20, s56, 1.0
; GFX9-NEXT:    v_add_f32_e64 v19, s55, 1.0
; GFX9-NEXT:    v_add_f32_e64 v18, s54, 1.0
; GFX9-NEXT:    v_add_f32_e64 v17, s53, 1.0
; GFX9-NEXT:    v_add_f32_e64 v16, s52, 1.0
; GFX9-NEXT:    v_add_f32_e64 v15, s51, 1.0
; GFX9-NEXT:    v_add_f32_e64 v14, s50, 1.0
; GFX9-NEXT:    v_add_f32_e64 v13, s49, 1.0
; GFX9-NEXT:    v_add_f32_e64 v12, s48, 1.0
; GFX9-NEXT:    v_add_f32_e64 v11, s47, 1.0
; GFX9-NEXT:    v_add_f32_e64 v10, s46, 1.0
; GFX9-NEXT:    v_add_f32_e64 v9, s45, 1.0
; GFX9-NEXT:    v_add_f32_e64 v8, s44, 1.0
; GFX9-NEXT:    v_add_f32_e64 v7, s43, 1.0
; GFX9-NEXT:    v_add_f32_e64 v6, s42, 1.0
; GFX9-NEXT:    v_add_f32_e64 v5, s41, 1.0
; GFX9-NEXT:    v_add_f32_e64 v4, s40, 1.0
; GFX9-NEXT:    v_add_f32_e64 v3, s39, 1.0
; GFX9-NEXT:    v_add_f32_e64 v2, s38, 1.0
; GFX9-NEXT:    v_add_f32_e64 v1, s37, 1.0
; GFX9-NEXT:    v_add_f32_e64 v0, s36, 1.0
; GFX9-NEXT:    s_branch .LBB33_5
; GFX9-NEXT:  .LBB33_3:
; GFX9-NEXT:    s_branch .LBB33_2
; GFX9-NEXT:  .LBB33_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB33_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v16f64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB33_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB33_4
; GFX11-NEXT:  .LBB33_2: ; %cmp.true
; GFX11-NEXT:    v_add_f32_e64 v31, s67, 1.0
; GFX11-NEXT:    v_add_f32_e64 v30, s66, 1.0
; GFX11-NEXT:    v_add_f32_e64 v29, s65, 1.0
; GFX11-NEXT:    v_add_f32_e64 v28, s64, 1.0
; GFX11-NEXT:    v_add_f32_e64 v27, s63, 1.0
; GFX11-NEXT:    v_add_f32_e64 v26, s62, 1.0
; GFX11-NEXT:    v_add_f32_e64 v25, s61, 1.0
; GFX11-NEXT:    v_add_f32_e64 v24, s60, 1.0
; GFX11-NEXT:    v_add_f32_e64 v23, s59, 1.0
; GFX11-NEXT:    v_add_f32_e64 v22, s58, 1.0
; GFX11-NEXT:    v_add_f32_e64 v21, s57, 1.0
; GFX11-NEXT:    v_add_f32_e64 v20, s56, 1.0
; GFX11-NEXT:    v_add_f32_e64 v19, s55, 1.0
; GFX11-NEXT:    v_add_f32_e64 v18, s54, 1.0
; GFX11-NEXT:    v_add_f32_e64 v17, s53, 1.0
; GFX11-NEXT:    v_add_f32_e64 v16, s52, 1.0
; GFX11-NEXT:    v_add_f32_e64 v15, s51, 1.0
; GFX11-NEXT:    v_add_f32_e64 v14, s50, 1.0
; GFX11-NEXT:    v_add_f32_e64 v13, s49, 1.0
; GFX11-NEXT:    v_add_f32_e64 v12, s48, 1.0
; GFX11-NEXT:    v_add_f32_e64 v11, s47, 1.0
; GFX11-NEXT:    v_add_f32_e64 v10, s46, 1.0
; GFX11-NEXT:    v_add_f32_e64 v9, s45, 1.0
; GFX11-NEXT:    v_add_f32_e64 v8, s44, 1.0
; GFX11-NEXT:    v_add_f32_e64 v7, s43, 1.0
; GFX11-NEXT:    v_add_f32_e64 v6, s42, 1.0
; GFX11-NEXT:    v_add_f32_e64 v5, s41, 1.0
; GFX11-NEXT:    v_add_f32_e64 v4, s40, 1.0
; GFX11-NEXT:    v_add_f32_e64 v3, s39, 1.0
; GFX11-NEXT:    v_add_f32_e64 v2, s38, 1.0
; GFX11-NEXT:    v_add_f32_e64 v1, s37, 1.0
; GFX11-NEXT:    v_add_f32_e64 v0, s36, 1.0
; GFX11-NEXT:    s_branch .LBB33_5
; GFX11-NEXT:  .LBB33_3:
; GFX11-NEXT:    s_branch .LBB33_2
; GFX11-NEXT:  .LBB33_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB33_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define <32 x float> @bitcast_v16f64_to_v32f32(<16 x double> %a, i32 %b) {
; SI-LABEL: bitcast_v16f64_to_v32f32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB34_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; SI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; SI-NEXT:  .LBB34_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v32f32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB34_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; VI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; VI-NEXT:  .LBB34_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v32f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB34_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX9-NEXT:  .LBB34_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v32f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB34_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX11-NEXT:  .LBB34_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define inreg <32 x float> @bitcast_v16f64_to_v32f32_scalar(<16 x double> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16f64_to_v32f32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s36, 0
; SI-NEXT:    v_writelane_b32 v32, s37, 1
; SI-NEXT:    v_writelane_b32 v32, s38, 2
; SI-NEXT:    v_writelane_b32 v32, s39, 3
; SI-NEXT:    v_writelane_b32 v32, s48, 4
; SI-NEXT:    v_writelane_b32 v32, s49, 5
; SI-NEXT:    v_writelane_b32 v32, s50, 6
; SI-NEXT:    v_writelane_b32 v32, s51, 7
; SI-NEXT:    v_writelane_b32 v32, s52, 8
; SI-NEXT:    v_writelane_b32 v32, s53, 9
; SI-NEXT:    v_writelane_b32 v32, s54, 10
; SI-NEXT:    v_writelane_b32 v32, s55, 11
; SI-NEXT:    v_writelane_b32 v32, s64, 12
; SI-NEXT:    v_writelane_b32 v32, s65, 13
; SI-NEXT:    v_writelane_b32 v32, s66, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v32, s67, 15
; SI-NEXT:    s_mov_b32 s49, s29
; SI-NEXT:    s_mov_b32 s48, s28
; SI-NEXT:    s_mov_b32 s47, s27
; SI-NEXT:    s_mov_b32 s46, s26
; SI-NEXT:    s_mov_b32 s45, s25
; SI-NEXT:    s_mov_b32 s44, s24
; SI-NEXT:    s_mov_b32 s43, s23
; SI-NEXT:    s_mov_b32 s42, s22
; SI-NEXT:    s_mov_b32 s41, s21
; SI-NEXT:    s_mov_b32 s40, s20
; SI-NEXT:    s_mov_b32 s39, s19
; SI-NEXT:    s_mov_b32 s38, s18
; SI-NEXT:    s_mov_b32 s37, s17
; SI-NEXT:    s_mov_b32 s36, s16
; SI-NEXT:    v_readfirstlane_b32 s67, v17
; SI-NEXT:    v_readfirstlane_b32 s66, v16
; SI-NEXT:    v_readfirstlane_b32 s65, v15
; SI-NEXT:    v_readfirstlane_b32 s64, v14
; SI-NEXT:    v_readfirstlane_b32 s63, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v12
; SI-NEXT:    v_readfirstlane_b32 s61, v11
; SI-NEXT:    v_readfirstlane_b32 s60, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s58, v8
; SI-NEXT:    v_readfirstlane_b32 s57, v7
; SI-NEXT:    v_readfirstlane_b32 s56, v6
; SI-NEXT:    v_readfirstlane_b32 s55, v5
; SI-NEXT:    v_readfirstlane_b32 s54, v4
; SI-NEXT:    v_readfirstlane_b32 s53, v3
; SI-NEXT:    v_readfirstlane_b32 s52, v2
; SI-NEXT:    v_readfirstlane_b32 s51, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s50, v0
; SI-NEXT:    s_cbranch_scc0 .LBB35_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB35_4
; SI-NEXT:  .LBB35_2: ; %cmp.true
; SI-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; SI-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; SI-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; SI-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; SI-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; SI-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; SI-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; SI-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; SI-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; SI-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; SI-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; SI-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; SI-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; SI-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; SI-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; SI-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; SI-NEXT:    s_branch .LBB35_5
; SI-NEXT:  .LBB35_3:
; SI-NEXT:    s_branch .LBB35_2
; SI-NEXT:  .LBB35_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB35_5: ; %end
; SI-NEXT:    v_readlane_b32 s67, v32, 15
; SI-NEXT:    v_readlane_b32 s66, v32, 14
; SI-NEXT:    v_readlane_b32 s65, v32, 13
; SI-NEXT:    v_readlane_b32 s64, v32, 12
; SI-NEXT:    v_readlane_b32 s55, v32, 11
; SI-NEXT:    v_readlane_b32 s54, v32, 10
; SI-NEXT:    v_readlane_b32 s53, v32, 9
; SI-NEXT:    v_readlane_b32 s52, v32, 8
; SI-NEXT:    v_readlane_b32 s51, v32, 7
; SI-NEXT:    v_readlane_b32 s50, v32, 6
; SI-NEXT:    v_readlane_b32 s49, v32, 5
; SI-NEXT:    v_readlane_b32 s48, v32, 4
; SI-NEXT:    v_readlane_b32 s39, v32, 3
; SI-NEXT:    v_readlane_b32 s38, v32, 2
; SI-NEXT:    v_readlane_b32 s37, v32, 1
; SI-NEXT:    v_readlane_b32 s36, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v32f32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s36, 0
; VI-NEXT:    v_writelane_b32 v32, s37, 1
; VI-NEXT:    v_writelane_b32 v32, s38, 2
; VI-NEXT:    v_writelane_b32 v32, s39, 3
; VI-NEXT:    v_writelane_b32 v32, s48, 4
; VI-NEXT:    v_writelane_b32 v32, s49, 5
; VI-NEXT:    v_writelane_b32 v32, s50, 6
; VI-NEXT:    v_writelane_b32 v32, s51, 7
; VI-NEXT:    v_writelane_b32 v32, s52, 8
; VI-NEXT:    v_writelane_b32 v32, s53, 9
; VI-NEXT:    v_writelane_b32 v32, s54, 10
; VI-NEXT:    v_writelane_b32 v32, s55, 11
; VI-NEXT:    v_writelane_b32 v32, s64, 12
; VI-NEXT:    v_writelane_b32 v32, s65, 13
; VI-NEXT:    v_writelane_b32 v32, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB35_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB35_4
; VI-NEXT:  .LBB35_2: ; %cmp.true
; VI-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; VI-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; VI-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; VI-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; VI-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; VI-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; VI-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; VI-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; VI-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; VI-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; VI-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; VI-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; VI-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; VI-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; VI-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; VI-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; VI-NEXT:    s_branch .LBB35_5
; VI-NEXT:  .LBB35_3:
; VI-NEXT:    s_branch .LBB35_2
; VI-NEXT:  .LBB35_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB35_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v32, 15
; VI-NEXT:    v_readlane_b32 s66, v32, 14
; VI-NEXT:    v_readlane_b32 s65, v32, 13
; VI-NEXT:    v_readlane_b32 s64, v32, 12
; VI-NEXT:    v_readlane_b32 s55, v32, 11
; VI-NEXT:    v_readlane_b32 s54, v32, 10
; VI-NEXT:    v_readlane_b32 s53, v32, 9
; VI-NEXT:    v_readlane_b32 s52, v32, 8
; VI-NEXT:    v_readlane_b32 s51, v32, 7
; VI-NEXT:    v_readlane_b32 s50, v32, 6
; VI-NEXT:    v_readlane_b32 s49, v32, 5
; VI-NEXT:    v_readlane_b32 s48, v32, 4
; VI-NEXT:    v_readlane_b32 s39, v32, 3
; VI-NEXT:    v_readlane_b32 s38, v32, 2
; VI-NEXT:    v_readlane_b32 s37, v32, 1
; VI-NEXT:    v_readlane_b32 s36, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v32f32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB35_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB35_4
; GFX9-NEXT:  .LBB35_2: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; GFX9-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; GFX9-NEXT:    s_branch .LBB35_5
; GFX9-NEXT:  .LBB35_3:
; GFX9-NEXT:    s_branch .LBB35_2
; GFX9-NEXT:  .LBB35_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB35_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v32f32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB35_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB35_4
; GFX11-NEXT:  .LBB35_2: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; GFX11-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; GFX11-NEXT:    s_branch .LBB35_5
; GFX11-NEXT:  .LBB35_3:
; GFX11-NEXT:    s_branch .LBB35_2
; GFX11-NEXT:  .LBB35_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB35_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) {
; SI-LABEL: bitcast_v32f32_to_v128i8:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB36_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v53, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v52, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB36_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB36_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_add_f32_e32 v32, 1.0, v32
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 16
; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v53, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v52, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB36_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v45
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v1, v1, v45
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v43
; SI-NEXT:    v_lshlrev_b32_e32 v41, 24, v41
; SI-NEXT:    v_or_b32_e32 v41, v41, v43
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v41
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v34
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v33
; SI-NEXT:    v_or_b32_e32 v2, v33, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v55
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v53
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v51
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v61
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v48
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v38
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v37
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v60
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v58
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v7
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v57
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v47
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v9
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v46
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v44
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v42
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v11
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v40
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v52
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v13
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v39
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v15
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v17
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v27
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v28
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v29
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v30
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v31
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v128i8:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB36_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v58, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v56, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v1
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:  .LBB36_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB36_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    v_add_f32_e32 v32, 1.0, v32
; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v58, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v56, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v1
; VI-NEXT:  .LBB36_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v55, v57, v41, s4
; VI-NEXT:    v_perm_b32 v1, v1, v48, s4
; VI-NEXT:    v_lshlrev_b32_e32 v48, 16, v55
; VI-NEXT:    v_or_b32_e32 v1, v1, v48
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v34, v56, s4
; VI-NEXT:    v_perm_b32 v2, v2, v53, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v47, v40, s4
; VI-NEXT:    v_perm_b32 v2, v3, v38, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v46, v45, s4
; VI-NEXT:    v_perm_b32 v1, v4, v52, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v33, v39, s4
; VI-NEXT:    v_perm_b32 v2, v5, v37, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v44, v43, s4
; VI-NEXT:    v_perm_b32 v1, v6, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v50, v54, s4
; VI-NEXT:    v_perm_b32 v2, v7, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v62, v61, s4
; VI-NEXT:    v_perm_b32 v1, v8, v36, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v9, v49, s4
; VI-NEXT:    v_perm_b32 v1, v60, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v59, v58, s4
; VI-NEXT:    v_perm_b32 v1, v10, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v11, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v12, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v13, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v14, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v15, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v16, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v17, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v18, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v19, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v20, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v21, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v22, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v23, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v24, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v25, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v26, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v27, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v28, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v29, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v30, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v31, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v32, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v42, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v128i8:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(39)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB36_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(40)
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v7
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v1
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; GFX9-NEXT:  .LBB36_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB36_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e32 v32, 1.0, v32
; GFX9-NEXT:    s_waitcnt vmcnt(38)
; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30
; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28
; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26
; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24
; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22
; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20
; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18
; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16
; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14
; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12
; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6
; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5
; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4
; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3
; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2
; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v7
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v1
; GFX9-NEXT:  .LBB36_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v11, v11, v36, s4
; GFX9-NEXT:    v_perm_b32 v12, v12, v35, s4
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v3, v3, v34, s4
; GFX9-NEXT:    v_perm_b32 v34, v38, v39, s4
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v4, v4, v51, s4
; GFX9-NEXT:    v_perm_b32 v6, v6, v33, s4
; GFX9-NEXT:    v_perm_b32 v33, v50, v54, s4
; GFX9-NEXT:    v_perm_b32 v55, v58, v41, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v59, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, v2, v52, s4
; GFX9-NEXT:    v_perm_b32 v37, v37, v53, s4
; GFX9-NEXT:    v_perm_b32 v52, v57, v40, s4
; GFX9-NEXT:    v_perm_b32 v5, v5, v56, s4
; GFX9-NEXT:    v_perm_b32 v7, v7, v47, s4
; GFX9-NEXT:    v_perm_b32 v8, v8, v46, s4
; GFX9-NEXT:    v_perm_b32 v9, v9, v45, s4
; GFX9-NEXT:    v_perm_b32 v10, v10, v44, s4
; GFX9-NEXT:    v_perm_b32 v13, v13, v60, s4
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v35, v36, v35, s4
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v38, v63, v38, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v15, v15, v39, s4
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v14, v14, v36, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v16, v16, v39, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v36, v36, v50, s4
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v39, v39, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v17, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v50, v50, v53, s4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX9-NEXT:    v_perm_b32 v1, v49, v43, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshl_or_b32 v1, v52, 16, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    v_perm_b32 v1, v48, v62, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshl_or_b32 v1, v34, 16, v5
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v61, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v6
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshl_or_b32 v1, v33, 16, v7
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v8
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshl_or_b32 v1, v37, 16, v9
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v10
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshl_or_b32 v1, v38, 16, v11
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v12
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshl_or_b32 v1, v35, 16, v13
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v14
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshl_or_b32 v1, v36, 16, v15
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v16
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshl_or_b32 v1, v39, 16, v17
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v18
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v19, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v50, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v20, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v21, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v22, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v23, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v24, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v25, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v26, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v27, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v28, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v30, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v31, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v32, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v42, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v32f32_to_v128i8:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x2
; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB36_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB36_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB36_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v20, 1.0, v20 :: v_dual_add_f32 v19, 1.0, v19
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v18, 1.0, v18 :: v_dual_add_f32 v17, 1.0, v17
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v24, 1.0, v24 :: v_dual_add_f32 v31, 1.0, v31
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v32, 1.0, v32 :: v_dual_add_f32 v23, 1.0, v23
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v15, 1.0, v15
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v26, 1.0, v26 :: v_dual_add_f32 v29, 1.0, v29
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v30, 1.0, v30 :: v_dual_add_f32 v25, 1.0, v25
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v13, 1.0, v13
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v28, 1.0, v28 :: v_dual_add_f32 v27, 1.0, v27
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v11, 1.0, v11
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 1.0, v10 :: v_dual_add_f32 v9, 1.0, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v7, 1.0, v7
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v22, 1.0, v22 :: v_dual_add_f32 v21, 1.0, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB36_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v1.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, v3.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v66, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v3, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, v4.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v65, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v6.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v39, 16, v2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v5.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, v8.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v7.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v66, v66, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v147, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v54, v64, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v39, 16, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v9.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v64, v65, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v66, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v55, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v10.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v11.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v64, 16, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v54, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v52, v55, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v12.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v13.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v14.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v15.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v12, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v13, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v14, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v15, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v16.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v17.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v18.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v19.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v16, v117, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v17, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v18, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v19, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v20.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v21.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v22.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v23.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v20, v103, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v21, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v99, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v14, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v24.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v15, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v25.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v26.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v27.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v27, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v28.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v29.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v30.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v31.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v28, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v29, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v30, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v31, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v32, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v29
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v32f32_to_v128i8:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x2
; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:12
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB36_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:  .LBB36_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB36_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v20, 1.0, v20 :: v_dual_add_f32 v19, 1.0, v19
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v18, 1.0, v18 :: v_dual_add_f32 v17, 1.0, v17
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v24, 1.0, v24 :: v_dual_add_f32 v31, 1.0, v31
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v32, 1.0, v32 :: v_dual_add_f32 v23, 1.0, v23
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v16, 1.0, v16 :: v_dual_add_f32 v15, 1.0, v15
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v26, 1.0, v26 :: v_dual_add_f32 v29, 1.0, v29
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v30, 1.0, v30 :: v_dual_add_f32 v25, 1.0, v25
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v14, 1.0, v14 :: v_dual_add_f32 v13, 1.0, v13
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v28, 1.0, v28 :: v_dual_add_f32 v27, 1.0, v27
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v6, 1.0, v6 :: v_dual_add_f32 v5, 1.0, v5
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v12, 1.0, v12 :: v_dual_add_f32 v11, 1.0, v11
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v4, 1.0, v4 :: v_dual_add_f32 v3, 1.0, v3
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v10, 1.0, v10 :: v_dual_add_f32 v9, 1.0, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v2, 1.0, v2 :: v_dual_add_f32 v1, 1.0, v1
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v8, 1.0, v8 :: v_dual_add_f32 v7, 1.0, v7
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v22, 1.0, v22 :: v_dual_add_f32 v21, 1.0, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:  .LBB36_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v74, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v73, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v72, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v61, v65, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v59, v58, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v43, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v55, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v57, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v56, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v64, v46, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v41, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v182, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v53, v180, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v177, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v55, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v64, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v65, 16, v8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v54, 16, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v13, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v164, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v14, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v162, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v15, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v151, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v16, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v149, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v17, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v146, v49, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v18, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v144, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v19, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v133, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v20, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v131, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v21, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v128, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v118, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v115, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v24, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v113, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v25, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v102, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v26, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v97, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v28, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v87, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v29, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v84, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v30, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v31, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v71, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v69, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:88
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define inreg <128 x i8> @bitcast_v32f32_to_v128i8_scalar(<32 x float> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32f32_to_v128i8_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_writelane_b32 v63, s70, 22
; SI-NEXT:    v_writelane_b32 v63, s71, 23
; SI-NEXT:    v_writelane_b32 v63, s80, 24
; SI-NEXT:    v_writelane_b32 v63, s81, 25
; SI-NEXT:    v_writelane_b32 v63, s82, 26
; SI-NEXT:    v_writelane_b32 v63, s83, 27
; SI-NEXT:    v_writelane_b32 v63, s84, 28
; SI-NEXT:    v_writelane_b32 v63, s85, 29
; SI-NEXT:    v_writelane_b32 v63, s86, 30
; SI-NEXT:    v_writelane_b32 v63, s87, 31
; SI-NEXT:    v_writelane_b32 v63, s96, 32
; SI-NEXT:    v_writelane_b32 v63, s97, 33
; SI-NEXT:    v_writelane_b32 v63, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s4, v19
; SI-NEXT:    v_writelane_b32 v63, s99, 35
; SI-NEXT:    v_readfirstlane_b32 s45, v18
; SI-NEXT:    v_readfirstlane_b32 s44, v17
; SI-NEXT:    v_readfirstlane_b32 s47, v16
; SI-NEXT:    v_readfirstlane_b32 s46, v15
; SI-NEXT:    v_readfirstlane_b32 s57, v14
; SI-NEXT:    v_readfirstlane_b32 s56, v13
; SI-NEXT:    v_readfirstlane_b32 s59, v12
; SI-NEXT:    v_readfirstlane_b32 s58, v11
; SI-NEXT:    v_readfirstlane_b32 s61, v10
; SI-NEXT:    v_readfirstlane_b32 s60, v9
; SI-NEXT:    v_readfirstlane_b32 s63, v8
; SI-NEXT:    v_readfirstlane_b32 s62, v7
; SI-NEXT:    v_readfirstlane_b32 s73, v6
; SI-NEXT:    v_readfirstlane_b32 s72, v5
; SI-NEXT:    v_readfirstlane_b32 s75, v4
; SI-NEXT:    v_readfirstlane_b32 s74, v3
; SI-NEXT:    v_readfirstlane_b32 s77, v2
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s76, v1
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr61 : SGPR spill to VGPR lane
; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB37_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s4, s45, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 17
; SI-NEXT:    s_lshr_b32 s4, s47, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 16
; SI-NEXT:    s_lshr_b32 s4, s57, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 15
; SI-NEXT:    s_lshr_b32 s4, s57, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 14
; SI-NEXT:    s_lshr_b32 s4, s59, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 13
; SI-NEXT:    s_lshr_b32 s4, s59, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 12
; SI-NEXT:    s_lshr_b32 s4, s59, 8
; SI-NEXT:    v_writelane_b32 v62, s4, 11
; SI-NEXT:    s_lshr_b32 s4, s61, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 10
; SI-NEXT:    s_lshr_b32 s4, s61, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 9
; SI-NEXT:    s_lshr_b32 s4, s61, 8
; SI-NEXT:    v_writelane_b32 v62, s4, 8
; SI-NEXT:    s_lshr_b32 s4, s63, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 7
; SI-NEXT:    s_lshr_b32 s4, s63, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 6
; SI-NEXT:    s_lshr_b32 s4, s63, 8
; SI-NEXT:    v_writelane_b32 v62, s4, 5
; SI-NEXT:    s_lshr_b32 s4, s73, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 4
; SI-NEXT:    s_lshr_b32 s4, s73, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 3
; SI-NEXT:    s_lshr_b32 s4, s73, 8
; SI-NEXT:    v_writelane_b32 v62, s4, 2
; SI-NEXT:    s_lshr_b32 s4, s75, 24
; SI-NEXT:    v_writelane_b32 v62, s4, 1
; SI-NEXT:    s_lshr_b32 s4, s75, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 0
; SI-NEXT:    s_lshr_b32 s4, s75, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 63
; SI-NEXT:    s_lshr_b32 s4, s77, 24
; SI-NEXT:    v_writelane_b32 v61, s4, 62
; SI-NEXT:    s_lshr_b32 s4, s77, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 61
; SI-NEXT:    s_lshr_b32 s4, s77, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 60
; SI-NEXT:    s_lshr_b32 s4, s29, 24
; SI-NEXT:    v_writelane_b32 v61, s4, 59
; SI-NEXT:    s_lshr_b32 s4, s29, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 58
; SI-NEXT:    s_lshr_b32 s4, s29, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 57
; SI-NEXT:    s_lshr_b32 s4, s27, 24
; SI-NEXT:    v_writelane_b32 v61, s4, 56
; SI-NEXT:    s_lshr_b32 s4, s27, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 55
; SI-NEXT:    s_lshr_b32 s4, s27, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 54
; SI-NEXT:    s_lshr_b32 s4, s25, 24
; SI-NEXT:    v_writelane_b32 v61, s4, 53
; SI-NEXT:    s_lshr_b32 s4, s25, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 52
; SI-NEXT:    s_lshr_b32 s4, s25, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 51
; SI-NEXT:    s_lshr_b32 s4, s23, 24
; SI-NEXT:    v_writelane_b32 v61, s4, 50
; SI-NEXT:    s_lshr_b32 s4, s23, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 49
; SI-NEXT:    s_lshr_b32 s4, s23, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 48
; SI-NEXT:    s_lshr_b32 s4, s21, 24
; SI-NEXT:    v_writelane_b32 v61, s4, 47
; SI-NEXT:    s_lshr_b32 s4, s21, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 46
; SI-NEXT:    s_lshr_b32 s4, s21, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 45
; SI-NEXT:    s_lshr_b32 s4, s19, 24
; SI-NEXT:    v_writelane_b32 v61, s4, 44
; SI-NEXT:    s_lshr_b32 s4, s19, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 43
; SI-NEXT:    s_lshr_b32 s4, s19, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 42
; SI-NEXT:    s_lshr_b32 s4, s17, 16
; SI-NEXT:    v_writelane_b32 v61, s4, 41
; SI-NEXT:    s_lshr_b32 s4, s17, 8
; SI-NEXT:    v_writelane_b32 v61, s4, 40
; SI-NEXT:    s_lshr_b64 s[4:5], s[44:45], 24
; SI-NEXT:    v_writelane_b32 v61, s4, 38
; SI-NEXT:    v_writelane_b32 v61, s5, 39
; SI-NEXT:    s_lshr_b64 s[4:5], s[44:45], 16
; SI-NEXT:    v_writelane_b32 v61, s4, 36
; SI-NEXT:    v_writelane_b32 v61, s5, 37
; SI-NEXT:    s_lshr_b64 s[4:5], s[44:45], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 34
; SI-NEXT:    v_writelane_b32 v61, s5, 35
; SI-NEXT:    s_lshr_b64 s[4:5], s[46:47], 24
; SI-NEXT:    v_writelane_b32 v61, s4, 32
; SI-NEXT:    v_writelane_b32 v61, s5, 33
; SI-NEXT:    s_lshr_b64 s[4:5], s[46:47], 16
; SI-NEXT:    v_writelane_b32 v61, s4, 30
; SI-NEXT:    v_writelane_b32 v61, s5, 31
; SI-NEXT:    s_lshr_b64 s[4:5], s[46:47], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 28
; SI-NEXT:    v_writelane_b32 v61, s5, 29
; SI-NEXT:    s_lshr_b64 s[4:5], s[56:57], 24
; SI-NEXT:    v_writelane_b32 v61, s4, 26
; SI-NEXT:    v_writelane_b32 v61, s5, 27
; SI-NEXT:    s_lshr_b64 s[4:5], s[56:57], 16
; SI-NEXT:    v_writelane_b32 v61, s4, 24
; SI-NEXT:    v_writelane_b32 v61, s5, 25
; SI-NEXT:    s_lshr_b64 s[4:5], s[56:57], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 22
; SI-NEXT:    v_writelane_b32 v61, s5, 23
; SI-NEXT:    s_lshr_b64 s[4:5], s[58:59], 24
; SI-NEXT:    v_writelane_b32 v61, s4, 20
; SI-NEXT:    v_writelane_b32 v61, s5, 21
; SI-NEXT:    s_lshr_b64 s[4:5], s[58:59], 16
; SI-NEXT:    v_writelane_b32 v61, s4, 18
; SI-NEXT:    v_writelane_b32 v61, s5, 19
; SI-NEXT:    s_lshr_b64 s[4:5], s[58:59], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 16
; SI-NEXT:    v_writelane_b32 v61, s5, 17
; SI-NEXT:    s_lshr_b64 s[4:5], s[60:61], 24
; SI-NEXT:    v_writelane_b32 v61, s4, 14
; SI-NEXT:    v_writelane_b32 v61, s5, 15
; SI-NEXT:    s_lshr_b64 s[4:5], s[60:61], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 12
; SI-NEXT:    v_writelane_b32 v61, s5, 13
; SI-NEXT:    s_lshr_b64 s[4:5], s[62:63], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 10
; SI-NEXT:    v_writelane_b32 v61, s5, 11
; SI-NEXT:    s_lshr_b64 s[4:5], s[72:73], 24
; SI-NEXT:    v_writelane_b32 v61, s4, 8
; SI-NEXT:    v_writelane_b32 v61, s5, 9
; SI-NEXT:    s_lshr_b64 s[4:5], s[72:73], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 6
; SI-NEXT:    v_writelane_b32 v61, s5, 7
; SI-NEXT:    s_lshr_b64 s[4:5], s[74:75], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 4
; SI-NEXT:    v_writelane_b32 v61, s5, 5
; SI-NEXT:    s_lshr_b64 s[4:5], s[76:77], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 2
; SI-NEXT:    v_writelane_b32 v61, s5, 3
; SI-NEXT:    s_lshr_b64 s[4:5], s[20:21], 8
; SI-NEXT:    v_writelane_b32 v61, s4, 0
; SI-NEXT:    s_lshr_b32 s79, s45, 16
; SI-NEXT:    s_lshr_b32 s89, s45, 8
; SI-NEXT:    s_lshr_b32 s91, s47, 16
; SI-NEXT:    s_lshr_b32 s78, s47, 8
; SI-NEXT:    s_lshr_b32 s88, s57, 8
; SI-NEXT:    s_lshr_b32 s90, s17, 24
; SI-NEXT:    s_lshr_b64 s[92:93], s[60:61], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[62:63], 24
; SI-NEXT:    s_lshr_b64 s[30:31], s[62:63], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[72:73], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[74:75], 24
; SI-NEXT:    s_lshr_b64 s[38:39], s[74:75], 16
; SI-NEXT:    s_lshr_b64 s[48:49], s[76:77], 24
; SI-NEXT:    s_lshr_b64 s[50:51], s[76:77], 16
; SI-NEXT:    s_lshr_b64 s[52:53], s[28:29], 24
; SI-NEXT:    s_lshr_b64 s[54:55], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[64:65], s[28:29], 8
; SI-NEXT:    s_lshr_b64 s[66:67], s[26:27], 24
; SI-NEXT:    s_lshr_b64 s[68:69], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[70:71], s[26:27], 8
; SI-NEXT:    s_lshr_b64 s[80:81], s[24:25], 24
; SI-NEXT:    s_lshr_b64 s[82:83], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[84:85], s[24:25], 8
; SI-NEXT:    s_lshr_b64 s[86:87], s[22:23], 24
; SI-NEXT:    s_lshr_b64 s[96:97], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[98:99], s[22:23], 8
; SI-NEXT:    s_lshr_b64 s[40:41], s[20:21], 24
; SI-NEXT:    s_lshr_b64 s[42:43], s[20:21], 16
; SI-NEXT:    v_writelane_b32 v61, s5, 1
; SI-NEXT:    s_lshr_b64 s[8:9], s[18:19], 24
; SI-NEXT:    s_lshr_b64 s[10:11], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[14:15], s[18:19], 8
; SI-NEXT:    s_lshr_b64 s[12:13], s[16:17], 24
; SI-NEXT:    s_lshr_b64 s[6:7], s[16:17], 16
; SI-NEXT:    s_lshr_b64 s[4:5], s[16:17], 8
; SI-NEXT:    s_cbranch_execnz .LBB37_4
; SI-NEXT:  .LBB37_2: ; %cmp.true
; SI-NEXT:    v_add_f32_e64 v1, s60, 1.0
; SI-NEXT:    v_add_f32_e64 v22, s26, 1.0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s26, v1
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_f32_e64 v1, s58, 1.0
; SI-NEXT:    v_add_f32_e64 v20, s28, 1.0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s28, v1
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_f32_e64 v1, s56, 1.0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s40, v1
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_f32_e64 v1, s46, 1.0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s42, v1
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_f32_e64 v1, s45, 1.0
; SI-NEXT:    v_add_f32_e64 v2, s44, 1.0
; SI-NEXT:    v_add_f32_e64 v13, s73, 1.0
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s44, v2
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 24, v1
; SI-NEXT:    v_add_f32_e64 v15, s75, 1.0
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 8, v13
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 24, v15
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v15
; SI-NEXT:    v_add_f32_e64 v17, s77, 1.0
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 8, v15
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 24, v17
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v17
; SI-NEXT:    v_add_f32_e64 v19, s29, 1.0
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 8, v17
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 24, v19
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
; SI-NEXT:    v_add_f32_e64 v21, s27, 1.0
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 8, v19
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 24, v21
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v21
; SI-NEXT:    v_add_f32_e64 v37, s17, 1.0
; SI-NEXT:    v_add_f32_e64 v38, s16, 1.0
; SI-NEXT:    v_add_f32_e64 v32, s19, 1.0
; SI-NEXT:    v_add_f32_e64 v35, s18, 1.0
; SI-NEXT:    v_add_f32_e64 v28, s21, 1.0
; SI-NEXT:    v_add_f32_e64 v31, s20, 1.0
; SI-NEXT:    v_add_f32_e64 v25, s23, 1.0
; SI-NEXT:    v_add_f32_e64 v26, s22, 1.0
; SI-NEXT:    v_add_f32_e64 v23, s25, 1.0
; SI-NEXT:    v_add_f32_e64 v24, s24, 1.0
; SI-NEXT:    v_add_f32_e64 v18, s76, 1.0
; SI-NEXT:    v_add_f32_e64 v16, s74, 1.0
; SI-NEXT:    v_add_f32_e64 v14, s72, 1.0
; SI-NEXT:    v_add_f32_e64 v11, s63, 1.0
; SI-NEXT:    v_add_f32_e64 v12, s62, 1.0
; SI-NEXT:    v_add_f32_e64 v9, s61, 1.0
; SI-NEXT:    v_add_f32_e64 v7, s59, 1.0
; SI-NEXT:    v_add_f32_e64 v5, s57, 1.0
; SI-NEXT:    v_add_f32_e64 v3, s47, 1.0
; SI-NEXT:    s_lshr_b64 s[46:47], s[44:45], 16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 8, v21
; SI-NEXT:    v_readfirstlane_b32 s16, v38
; SI-NEXT:    v_readfirstlane_b32 s17, v37
; SI-NEXT:    v_readfirstlane_b32 s12, v35
; SI-NEXT:    v_readfirstlane_b32 s13, v32
; SI-NEXT:    v_readfirstlane_b32 s6, v31
; SI-NEXT:    v_readfirstlane_b32 s7, v28
; SI-NEXT:    v_readfirstlane_b32 s4, v26
; SI-NEXT:    v_readfirstlane_b32 s5, v25
; SI-NEXT:    v_readfirstlane_b32 s8, v24
; SI-NEXT:    v_readfirstlane_b32 s9, v23
; SI-NEXT:    v_readfirstlane_b32 s10, v22
; SI-NEXT:    v_readfirstlane_b32 s11, v21
; SI-NEXT:    v_readfirstlane_b32 s14, v20
; SI-NEXT:    v_readfirstlane_b32 s15, v19
; SI-NEXT:    v_readfirstlane_b32 s18, v18
; SI-NEXT:    v_readfirstlane_b32 s19, v17
; SI-NEXT:    v_readfirstlane_b32 s20, v16
; SI-NEXT:    v_readfirstlane_b32 s21, v15
; SI-NEXT:    v_readfirstlane_b32 s22, v14
; SI-NEXT:    v_readfirstlane_b32 s23, v13
; SI-NEXT:    v_readfirstlane_b32 s24, v12
; SI-NEXT:    v_readfirstlane_b32 s25, v11
; SI-NEXT:    v_readfirstlane_b32 s27, v9
; SI-NEXT:    v_readfirstlane_b32 s29, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    s_lshr_b64 vcc, s[44:45], 24
; SI-NEXT:    v_writelane_b32 v61, s46, 36
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 24, v23
; SI-NEXT:    v_writelane_b32 v61, s47, 37
; SI-NEXT:    s_lshr_b64 s[44:45], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[58:59], s[42:43], 24
; SI-NEXT:    s_lshr_b64 s[60:61], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[42:43], 8
; SI-NEXT:    s_lshr_b64 s[72:73], s[40:41], 24
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[40:41], 8
; SI-NEXT:    s_lshr_b64 s[78:79], s[28:29], 24
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[28:29], s[28:29], 8
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; SI-NEXT:    s_lshr_b64 s[92:93], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[26:27], s[26:27], 8
; SI-NEXT:    s_lshr_b64 s[94:95], s[24:25], 24
; SI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[24:25], s[24:25], 8
; SI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; SI-NEXT:    s_lshr_b64 s[46:47], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[22:23], s[22:23], 8
; SI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; SI-NEXT:    s_lshr_b64 s[38:39], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[20:21], s[20:21], 8
; SI-NEXT:    s_lshr_b64 s[48:49], s[18:19], 24
; SI-NEXT:    s_lshr_b64 s[50:51], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[18:19], s[18:19], 8
; SI-NEXT:    s_lshr_b64 s[52:53], s[14:15], 24
; SI-NEXT:    s_lshr_b64 s[54:55], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[64:65], s[14:15], 8
; SI-NEXT:    s_lshr_b64 s[66:67], s[10:11], 24
; SI-NEXT:    s_lshr_b64 s[68:69], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[70:71], s[10:11], 8
; SI-NEXT:    s_lshr_b64 s[80:81], s[8:9], 24
; SI-NEXT:    s_lshr_b64 s[82:83], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[84:85], s[8:9], 8
; SI-NEXT:    s_lshr_b64 s[86:87], s[4:5], 24
; SI-NEXT:    s_lshr_b64 s[96:97], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[98:99], s[4:5], 8
; SI-NEXT:    s_lshr_b64 s[40:41], s[6:7], 24
; SI-NEXT:    s_lshr_b64 s[42:43], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[8:9], s[6:7], 8
; SI-NEXT:    s_lshr_b64 s[56:57], s[12:13], 24
; SI-NEXT:    s_lshr_b64 s[10:11], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[14:15], s[12:13], 8
; SI-NEXT:    s_lshr_b64 s[12:13], s[16:17], 24
; SI-NEXT:    s_lshr_b64 s[6:7], s[16:17], 16
; SI-NEXT:    s_lshr_b64 s[4:5], s[16:17], 8
; SI-NEXT:    s_mov_b32 s16, vcc_lo
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v1
; SI-NEXT:    v_lshrrev_b32_e32 v41, 8, v1
; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v3
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v44, 8, v3
; SI-NEXT:    v_lshrrev_b32_e32 v45, 24, v5
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v47, 8, v5
; SI-NEXT:    v_lshrrev_b32_e32 v56, 24, v7
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v58, 8, v7
; SI-NEXT:    v_lshrrev_b32_e32 v59, 24, v9
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v27, 8, v9
; SI-NEXT:    v_lshrrev_b32_e32 v29, 24, v11
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; SI-NEXT:    v_lshrrev_b32_e32 v34, 24, v13
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v13
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v39, 8, v23
; SI-NEXT:    v_lshrrev_b32_e32 v48, 24, v25
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v4, 8, v25
; SI-NEXT:    v_lshrrev_b32_e32 v6, 24, v28
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v51, 8, v28
; SI-NEXT:    v_lshrrev_b32_e32 v52, 24, v32
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v32
; SI-NEXT:    v_lshrrev_b32_e32 v53, 8, v32
; SI-NEXT:    v_lshrrev_b32_e32 v10, 24, v37
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v37
; SI-NEXT:    v_lshrrev_b32_e32 v55, 8, v37
; SI-NEXT:    s_branch .LBB37_5
; SI-NEXT:  .LBB37_3:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s8, 0
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s9, 1
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 2
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s91, 3
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s91, 5
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 6
; SI-NEXT:    v_writelane_b32 v61, s91, 7
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 8
; SI-NEXT:    v_writelane_b32 v61, s91, 9
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 10
; SI-NEXT:    v_writelane_b32 v61, s91, 11
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 12
; SI-NEXT:    v_writelane_b32 v61, s91, 13
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 14
; SI-NEXT:    v_writelane_b32 v61, s91, 15
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 16
; SI-NEXT:    v_writelane_b32 v61, s91, 17
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 18
; SI-NEXT:    v_writelane_b32 v61, s91, 19
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 20
; SI-NEXT:    v_writelane_b32 v61, s91, 21
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 22
; SI-NEXT:    v_writelane_b32 v61, s91, 23
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 24
; SI-NEXT:    v_writelane_b32 v61, s91, 25
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 26
; SI-NEXT:    v_writelane_b32 v61, s91, 27
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 28
; SI-NEXT:    v_writelane_b32 v61, s91, 29
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 30
; SI-NEXT:    v_writelane_b32 v61, s91, 31
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 32
; SI-NEXT:    v_writelane_b32 v61, s91, 33
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 34
; SI-NEXT:    v_writelane_b32 v61, s91, 35
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 36
; SI-NEXT:    v_writelane_b32 v61, s91, 37
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s90, 38
; SI-NEXT:    v_writelane_b32 v61, s91, 39
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    s_branch .LBB37_2
; SI-NEXT:  .LBB37_4:
; SI-NEXT:    v_mov_b32_e32 v1, s16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_readlane_b32 s5, v61, 40
; SI-NEXT:    v_mov_b32_e32 v55, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 41
; SI-NEXT:    v_mov_b32_e32 v54, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 42
; SI-NEXT:    v_mov_b32_e32 v53, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 43
; SI-NEXT:    v_mov_b32_e32 v8, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 44
; SI-NEXT:    v_mov_b32_e32 v52, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 45
; SI-NEXT:    v_mov_b32_e32 v51, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 46
; SI-NEXT:    v_mov_b32_e32 v50, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 47
; SI-NEXT:    v_mov_b32_e32 v6, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 48
; SI-NEXT:    v_mov_b32_e32 v4, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 49
; SI-NEXT:    v_mov_b32_e32 v49, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 50
; SI-NEXT:    v_mov_b32_e32 v48, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 51
; SI-NEXT:    v_mov_b32_e32 v39, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 52
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 53
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 54
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 55
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 56
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 57
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 58
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 59
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 60
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 61
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 62
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v61, 63
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 0
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 1
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 2
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 3
; SI-NEXT:    v_mov_b32_e32 v36, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 4
; SI-NEXT:    v_mov_b32_e32 v34, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 5
; SI-NEXT:    v_mov_b32_e32 v33, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 6
; SI-NEXT:    v_mov_b32_e32 v30, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 7
; SI-NEXT:    v_mov_b32_e32 v29, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 8
; SI-NEXT:    v_mov_b32_e32 v27, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 9
; SI-NEXT:    v_mov_b32_e32 v60, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 10
; SI-NEXT:    v_mov_b32_e32 v59, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 11
; SI-NEXT:    v_mov_b32_e32 v58, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 12
; SI-NEXT:    v_mov_b32_e32 v57, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 13
; SI-NEXT:    v_mov_b32_e32 v1, s60
; SI-NEXT:    v_mov_b32_e32 v56, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 14
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, s58
; SI-NEXT:    v_mov_b32_e32 v46, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 15
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, s56
; SI-NEXT:    v_mov_b32_e32 v45, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, s46
; SI-NEXT:    v_mov_b32_e32 v42, s5
; SI-NEXT:    v_readlane_b32 s5, v62, 17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v3, s47
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, s44
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s5
; SI-NEXT:    v_readlane_b32 s46, v61, 38
; SI-NEXT:    v_mov_b32_e32 v32, s19
; SI-NEXT:    v_mov_b32_e32 v28, s21
; SI-NEXT:    v_mov_b32_e32 v25, s23
; SI-NEXT:    v_mov_b32_e32 v23, s25
; SI-NEXT:    v_mov_b32_e32 v21, s27
; SI-NEXT:    v_mov_b32_e32 v19, s29
; SI-NEXT:    v_mov_b32_e32 v17, s77
; SI-NEXT:    v_mov_b32_e32 v15, s75
; SI-NEXT:    v_mov_b32_e32 v13, s73
; SI-NEXT:    v_mov_b32_e32 v11, s63
; SI-NEXT:    v_mov_b32_e32 v9, s61
; SI-NEXT:    v_mov_b32_e32 v7, s59
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, s45
; SI-NEXT:    v_mov_b32_e32 v10, s90
; SI-NEXT:    v_mov_b32_e32 v47, s88
; SI-NEXT:    v_mov_b32_e32 v44, s78
; SI-NEXT:    v_mov_b32_e32 v43, s91
; SI-NEXT:    v_mov_b32_e32 v41, s89
; SI-NEXT:    v_mov_b32_e32 v40, s79
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v12, s62
; SI-NEXT:    v_mov_b32_e32 v14, s72
; SI-NEXT:    v_mov_b32_e32 v16, s74
; SI-NEXT:    v_mov_b32_e32 v18, s76
; SI-NEXT:    v_mov_b32_e32 v20, s28
; SI-NEXT:    v_mov_b32_e32 v22, s26
; SI-NEXT:    v_mov_b32_e32 v24, s24
; SI-NEXT:    v_mov_b32_e32 v26, s22
; SI-NEXT:    v_mov_b32_e32 v31, s20
; SI-NEXT:    v_mov_b32_e32 v35, s18
; SI-NEXT:    s_mov_b32 s16, s46
; SI-NEXT:    v_readlane_b32 s44, v61, 34
; SI-NEXT:    v_readlane_b32 s58, v61, 32
; SI-NEXT:    v_readlane_b32 s60, v61, 30
; SI-NEXT:    v_readlane_b32 s62, v61, 28
; SI-NEXT:    v_readlane_b32 s72, v61, 26
; SI-NEXT:    v_readlane_b32 s74, v61, 24
; SI-NEXT:    v_readlane_b32 s76, v61, 22
; SI-NEXT:    v_readlane_b32 s78, v61, 20
; SI-NEXT:    v_readlane_b32 s88, v61, 18
; SI-NEXT:    v_readlane_b32 s28, v61, 16
; SI-NEXT:    v_readlane_b32 s90, v61, 14
; SI-NEXT:    v_readlane_b32 s26, v61, 12
; SI-NEXT:    v_readlane_b32 s24, v61, 10
; SI-NEXT:    v_readlane_b32 s22, v61, 6
; SI-NEXT:    s_mov_b32 s46, s34
; SI-NEXT:    v_readlane_b32 s34, v61, 8
; SI-NEXT:    v_readlane_b32 s20, v61, 4
; SI-NEXT:    v_readlane_b32 s18, v61, 2
; SI-NEXT:    s_mov_b32 s56, s8
; SI-NEXT:    v_readlane_b32 s8, v61, 0
; SI-NEXT:    v_mov_b32_e32 v37, s17
; SI-NEXT:    v_mov_b32_e32 v5, s57
; SI-NEXT:    v_readlane_b32 s47, v61, 39
; SI-NEXT:    v_readlane_b32 s45, v61, 35
; SI-NEXT:    v_readlane_b32 s59, v61, 33
; SI-NEXT:    v_readlane_b32 s61, v61, 31
; SI-NEXT:    v_readlane_b32 s63, v61, 29
; SI-NEXT:    v_readlane_b32 s73, v61, 27
; SI-NEXT:    v_readlane_b32 s75, v61, 25
; SI-NEXT:    v_readlane_b32 s77, v61, 23
; SI-NEXT:    v_readlane_b32 s79, v61, 21
; SI-NEXT:    v_readlane_b32 s89, v61, 19
; SI-NEXT:    v_readlane_b32 s29, v61, 17
; SI-NEXT:    v_readlane_b32 s91, v61, 15
; SI-NEXT:    v_readlane_b32 s27, v61, 13
; SI-NEXT:    v_readlane_b32 s25, v61, 11
; SI-NEXT:    v_readlane_b32 s23, v61, 7
; SI-NEXT:    v_readlane_b32 s35, v61, 9
; SI-NEXT:    v_readlane_b32 s21, v61, 5
; SI-NEXT:    v_readlane_b32 s19, v61, 3
; SI-NEXT:    v_readlane_b32 s9, v61, 1
; SI-NEXT:  .LBB37_5: ; %end
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v38, 0xff, v38
; SI-NEXT:    s_lshl_b32 s5, s4, 8
; SI-NEXT:    v_or_b32_e32 v38, s5, v38
; SI-NEXT:    s_and_b32 s5, s6, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s7, s12, 24
; SI-NEXT:    v_and_b32_e32 v38, 0xffff, v38
; SI-NEXT:    s_or_b32 s5, s7, s5
; SI-NEXT:    v_or_b32_e32 v38, s5, v38
; SI-NEXT:    buffer_store_dword v38, v0, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v37, 0xff, v37
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v55
; SI-NEXT:    v_or_b32_e32 v37, v37, v38
; SI-NEXT:    v_and_b32_e32 v38, 0xff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v38
; SI-NEXT:    v_and_b32_e32 v37, 0xffff, v37
; SI-NEXT:    v_or_b32_e32 v10, v37, v10
; SI-NEXT:    v_add_i32_e32 v37, vcc, 4, v0
; SI-NEXT:    buffer_store_dword v10, v37, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v35
; SI-NEXT:    s_lshl_b32 s5, s14, 8
; SI-NEXT:    v_or_b32_e32 v10, s5, v10
; SI-NEXT:    s_and_b32 s5, s10, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s7, s56, 24
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    s_or_b32 s5, s7, s5
; SI-NEXT:    v_or_b32_e32 v10, s5, v10
; SI-NEXT:    v_add_i32_e32 v35, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v10, v35, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 8, v53
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v10, v10, v32
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v52
; SI-NEXT:    v_or_b32_e32 v8, v32, v8
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v8, v10, v8
; SI-NEXT:    v_add_i32_e32 v10, vcc, 12, v0
; SI-NEXT:    buffer_store_dword v8, v10, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v31
; SI-NEXT:    s_lshl_b32 s5, s8, 8
; SI-NEXT:    s_and_b32 s4, s42, 0xff
; SI-NEXT:    v_or_b32_e32 v8, s5, v8
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s5, s40, 24
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v8, s4, v8
; SI-NEXT:    v_add_i32_e32 v10, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v8, v10, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v10, 8, v51
; SI-NEXT:    v_or_b32_e32 v8, v8, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v50
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v10
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v6, v8, v6
; SI-NEXT:    v_add_i32_e32 v8, vcc, 20, v0
; SI-NEXT:    buffer_store_dword v6, v8, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v26
; SI-NEXT:    s_lshl_b32 s4, s98, 8
; SI-NEXT:    v_or_b32_e32 v6, s4, v6
; SI-NEXT:    s_and_b32 s4, s96, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s5, s86, 24
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v6, s4, v6
; SI-NEXT:    v_add_i32_e32 v8, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v6, v8, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_lshlrev_b32_e32 v8, 24, v48
; SI-NEXT:    v_or_b32_e32 v6, v8, v6
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 28, v0
; SI-NEXT:    buffer_store_dword v4, v6, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v24
; SI-NEXT:    s_lshl_b32 s4, s84, 8
; SI-NEXT:    v_or_b32_e32 v4, s4, v4
; SI-NEXT:    s_and_b32 s4, s82, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s5, s80, 24
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v4, s4, v4
; SI-NEXT:    v_add_i32_e32 v6, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v4, v6, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v39
; SI-NEXT:    v_or_b32_e32 v4, v4, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    s_lshl_b32 s4, s70, 8
; SI-NEXT:    s_lshl_b32 s5, s66, 24
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_readlane_b32 s99, v63, 35
; SI-NEXT:    v_readlane_b32 s98, v63, 34
; SI-NEXT:    v_readlane_b32 s97, v63, 33
; SI-NEXT:    v_readlane_b32 s96, v63, 32
; SI-NEXT:    v_readlane_b32 s87, v63, 31
; SI-NEXT:    v_readlane_b32 s86, v63, 30
; SI-NEXT:    v_readlane_b32 s85, v63, 29
; SI-NEXT:    v_readlane_b32 s84, v63, 28
; SI-NEXT:    v_readlane_b32 s83, v63, 27
; SI-NEXT:    v_readlane_b32 s82, v63, 26
; SI-NEXT:    v_readlane_b32 s81, v63, 25
; SI-NEXT:    v_readlane_b32 s80, v63, 24
; SI-NEXT:    v_readlane_b32 s71, v63, 23
; SI-NEXT:    v_readlane_b32 s70, v63, 22
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    v_or_b32_e32 v2, v6, v2
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    v_add_i32_e32 v4, vcc, 36, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v22
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s68, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    v_add_i32_e32 v4, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v21
; SI-NEXT:    s_lshl_b32 s4, s64, 8
; SI-NEXT:    s_lshl_b32 s5, s52, 24
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 44, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v20
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s54, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    v_add_i32_e32 v4, vcc, 48, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v19
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s18, 8
; SI-NEXT:    s_lshl_b32 s5, s48, 24
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 52, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v18
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s50, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    v_add_i32_e32 v4, vcc, 56, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v17
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s20, 8
; SI-NEXT:    s_lshl_b32 s5, s36, 24
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 60, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v16
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s38, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    v_add_i32_e32 v4, vcc, 64, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v15
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s22, 8
; SI-NEXT:    s_lshl_b32 s5, s34, 24
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x44, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v14
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v34
; SI-NEXT:    s_lshl_b32 s4, s24, 8
; SI-NEXT:    s_lshl_b32 s5, s94, 24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x4c, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v12
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s30, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x50, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v33
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v29
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x54, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s26, 8
; SI-NEXT:    s_lshl_b32 s5, s90, 24
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x58, v0
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v59
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s92, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v27
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v60
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x5c, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s28, 8
; SI-NEXT:    s_lshl_b32 s5, s78, 24
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x60, v0
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v56
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s88, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v58
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x64, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s76, 8
; SI-NEXT:    s_lshl_b32 s5, s72, 24
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x68, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s74, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v47
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v46
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v45
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x6c, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s62, 8
; SI-NEXT:    s_lshl_b32 s5, s58, 24
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x70, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    s_and_b32 s4, s60, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v44
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v43
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v42
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x74, v0
; SI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    s_lshl_b32 s4, s44, 8
; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x78, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    v_readlane_b32 s4, v61, 36
; SI-NEXT:    v_readlane_b32 s5, v61, 37
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s5, s16, 24
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_or_b32_e32 v2, s4, v2
; SI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v41
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v40
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v128i8_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v63, s30, 0
; VI-NEXT:    v_writelane_b32 v63, s31, 1
; VI-NEXT:    v_writelane_b32 v63, s34, 2
; VI-NEXT:    v_writelane_b32 v63, s35, 3
; VI-NEXT:    v_writelane_b32 v63, s36, 4
; VI-NEXT:    v_writelane_b32 v63, s37, 5
; VI-NEXT:    v_writelane_b32 v63, s38, 6
; VI-NEXT:    v_writelane_b32 v63, s39, 7
; VI-NEXT:    v_writelane_b32 v63, s48, 8
; VI-NEXT:    v_writelane_b32 v63, s49, 9
; VI-NEXT:    v_writelane_b32 v63, s50, 10
; VI-NEXT:    v_writelane_b32 v63, s51, 11
; VI-NEXT:    v_writelane_b32 v63, s52, 12
; VI-NEXT:    v_writelane_b32 v63, s53, 13
; VI-NEXT:    v_writelane_b32 v63, s54, 14
; VI-NEXT:    v_writelane_b32 v63, s55, 15
; VI-NEXT:    v_writelane_b32 v63, s64, 16
; VI-NEXT:    v_writelane_b32 v63, s65, 17
; VI-NEXT:    v_writelane_b32 v63, s66, 18
; VI-NEXT:    v_writelane_b32 v63, s67, 19
; VI-NEXT:    v_writelane_b32 v63, s68, 20
; VI-NEXT:    v_writelane_b32 v63, s69, 21
; VI-NEXT:    v_writelane_b32 v63, s70, 22
; VI-NEXT:    v_writelane_b32 v63, s71, 23
; VI-NEXT:    v_writelane_b32 v63, s80, 24
; VI-NEXT:    v_writelane_b32 v63, s81, 25
; VI-NEXT:    v_writelane_b32 v63, s82, 26
; VI-NEXT:    v_writelane_b32 v63, s83, 27
; VI-NEXT:    v_writelane_b32 v63, s84, 28
; VI-NEXT:    v_writelane_b32 v63, s85, 29
; VI-NEXT:    v_writelane_b32 v63, s86, 30
; VI-NEXT:    v_readfirstlane_b32 s44, v19
; VI-NEXT:    v_writelane_b32 v63, s87, 31
; VI-NEXT:    v_readfirstlane_b32 s5, v18
; VI-NEXT:    v_readfirstlane_b32 s4, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s6, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s8, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s10, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s12, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s14, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s40, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s42, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    s_cmp_lg_u32 s44, 0
; VI-NEXT:    v_readfirstlane_b32 s44, v1
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; VI-NEXT:    s_cbranch_scc0 .LBB37_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 57
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s14, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s14, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s41, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s41, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s40, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s40, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s43, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s43, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s42, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s42, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s29, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s28, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s27, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s26, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 1
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    s_lshr_b32 s80, s25, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 0
; VI-NEXT:    s_lshr_b32 s81, s24, 8
; VI-NEXT:    s_lshr_b32 s82, s23, 24
; VI-NEXT:    s_lshr_b32 s83, s23, 16
; VI-NEXT:    s_lshr_b32 s85, s23, 8
; VI-NEXT:    s_lshr_b32 s84, s22, 16
; VI-NEXT:    s_lshr_b32 s86, s22, 8
; VI-NEXT:    s_lshr_b32 s87, s21, 24
; VI-NEXT:    s_lshr_b32 s50, s21, 16
; VI-NEXT:    s_lshr_b32 s52, s21, 8
; VI-NEXT:    s_lshr_b32 s51, s20, 16
; VI-NEXT:    s_lshr_b32 s53, s20, 8
; VI-NEXT:    s_lshr_b32 s54, s19, 24
; VI-NEXT:    s_lshr_b32 s55, s19, 16
; VI-NEXT:    s_lshr_b32 s65, s19, 8
; VI-NEXT:    s_lshr_b32 s64, s18, 16
; VI-NEXT:    s_lshr_b32 s66, s18, 8
; VI-NEXT:    s_lshr_b32 s67, s17, 24
; VI-NEXT:    s_lshr_b32 s68, s17, 16
; VI-NEXT:    s_lshr_b32 s70, s17, 8
; VI-NEXT:    s_lshr_b32 s69, s16, 16
; VI-NEXT:    s_lshr_b32 s71, s16, 8
; VI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:    s_cbranch_execnz .LBB37_4
; VI-NEXT:  .LBB37_2: ; %cmp.true
; VI-NEXT:    v_add_f32_e64 v2, s5, 1.0
; VI-NEXT:    v_add_f32_e64 v1, s4, 1.0
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[1:2]
; VI-NEXT:    v_add_f32_e64 v4, s7, 1.0
; VI-NEXT:    v_add_f32_e64 v3, s6, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[3:4]
; VI-NEXT:    v_add_f32_e64 v6, s9, 1.0
; VI-NEXT:    v_add_f32_e64 v5, s8, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[5:6]
; VI-NEXT:    v_add_f32_e64 v8, s11, 1.0
; VI-NEXT:    v_add_f32_e64 v7, s10, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[7:8]
; VI-NEXT:    v_add_f32_e64 v10, s13, 1.0
; VI-NEXT:    v_add_f32_e64 v9, s12, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    v_add_f32_e64 v12, s15, 1.0
; VI-NEXT:    v_add_f32_e64 v11, s14, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    v_add_f32_e64 v14, s41, 1.0
; VI-NEXT:    v_add_f32_e64 v13, s40, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    v_add_f32_e64 v16, s43, 1.0
; VI-NEXT:    v_add_f32_e64 v15, s42, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    v_add_f32_e64 v18, s45, 1.0
; VI-NEXT:    v_add_f32_e64 v17, s44, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; VI-NEXT:    v_add_f32_e64 v20, s29, 1.0
; VI-NEXT:    v_add_f32_e64 v19, s28, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    v_add_f32_e64 v22, s27, 1.0
; VI-NEXT:    v_add_f32_e64 v21, s26, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v2
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v4
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v4
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v4
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v3
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v5
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v8
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v7
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    v_add_f32_e64 v28, s21, 1.0
; VI-NEXT:    v_add_f32_e64 v27, s20, 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    v_add_f32_e64 v30, s19, 1.0
; VI-NEXT:    v_add_f32_e64 v29, s18, 1.0
; VI-NEXT:    v_add_f32_e64 v24, s25, 1.0
; VI-NEXT:    v_add_f32_e64 v23, s24, 1.0
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    v_add_f32_e64 v32, s17, 1.0
; VI-NEXT:    v_add_f32_e64 v31, s16, 1.0
; VI-NEXT:    v_add_f32_e64 v26, s23, 1.0
; VI-NEXT:    v_add_f32_e64 v25, s22, 1.0
; VI-NEXT:    v_lshrrev_b64 v[53:54], 24, v[23:24]
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[25:26]
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    v_lshrrev_b32_e32 v35, 24, v24
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v37, 16, v23
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v23
; VI-NEXT:    v_lshrrev_b32_e32 v55, 24, v26
; VI-NEXT:    v_lshrrev_b32_e32 v38, 16, v26
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v26
; VI-NEXT:    v_lshrrev_b32_e32 v42, 16, v25
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v25
; VI-NEXT:    v_lshrrev_b32_e32 v44, 24, v28
; VI-NEXT:    v_lshrrev_b32_e32 v43, 16, v28
; VI-NEXT:    v_lshrrev_b32_e32 v45, 8, v28
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v27
; VI-NEXT:    v_lshrrev_b32_e32 v46, 8, v27
; VI-NEXT:    v_lshrrev_b32_e32 v57, 24, v30
; VI-NEXT:    v_lshrrev_b32_e32 v56, 16, v30
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v30
; VI-NEXT:    v_lshrrev_b32_e32 v48, 16, v29
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v29
; VI-NEXT:    v_lshrrev_b32_e32 v59, 24, v32
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v32
; VI-NEXT:    v_lshrrev_b32_e32 v60, 8, v32
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    v_lshrrev_b32_e32 v61, 8, v31
; VI-NEXT:    s_branch .LBB37_5
; VI-NEXT:  .LBB37_3:
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr71
; VI-NEXT:    ; implicit-def: $sgpr69
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr68
; VI-NEXT:    ; implicit-def: $sgpr67
; VI-NEXT:    ; implicit-def: $sgpr66
; VI-NEXT:    ; implicit-def: $sgpr64
; VI-NEXT:    ; implicit-def: $sgpr65
; VI-NEXT:    ; implicit-def: $sgpr55
; VI-NEXT:    ; implicit-def: $sgpr54
; VI-NEXT:    ; implicit-def: $sgpr53
; VI-NEXT:    ; implicit-def: $sgpr51
; VI-NEXT:    ; implicit-def: $sgpr52
; VI-NEXT:    ; implicit-def: $sgpr50
; VI-NEXT:    ; implicit-def: $sgpr87
; VI-NEXT:    ; implicit-def: $sgpr86
; VI-NEXT:    ; implicit-def: $sgpr84
; VI-NEXT:    ; implicit-def: $sgpr85
; VI-NEXT:    ; implicit-def: $sgpr83
; VI-NEXT:    ; implicit-def: $sgpr82
; VI-NEXT:    ; implicit-def: $sgpr81
; VI-NEXT:    ; implicit-def: $sgpr80
; VI-NEXT:    ; implicit-def: $sgpr48
; VI-NEXT:    ; implicit-def: $sgpr38
; VI-NEXT:    ; implicit-def: $sgpr36
; VI-NEXT:    ; implicit-def: $sgpr34
; VI-NEXT:    ; implicit-def: $sgpr30
; VI-NEXT:    ; implicit-def: $sgpr90
; VI-NEXT:    ; implicit-def: $sgpr88
; VI-NEXT:    ; implicit-def: $sgpr78
; VI-NEXT:    ; implicit-def: $sgpr76
; VI-NEXT:    ; implicit-def: $sgpr74
; VI-NEXT:    ; implicit-def: $sgpr72
; VI-NEXT:    ; implicit-def: $sgpr62
; VI-NEXT:    ; implicit-def: $sgpr60
; VI-NEXT:    ; implicit-def: $sgpr58
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    s_branch .LBB37_2
; VI-NEXT:  .LBB37_4:
; VI-NEXT:    v_mov_b32_e32 v53, s46
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s56
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 0
; VI-NEXT:    v_mov_b32_e32 v37, s4
; VI-NEXT:    v_mov_b32_e32 v34, s80
; VI-NEXT:    v_readlane_b32 s4, v62, 1
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 2
; VI-NEXT:    v_mov_b32_e32 v35, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 3
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 4
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 5
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 6
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 7
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 8
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 9
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 10
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 11
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 12
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 13
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 14
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 15
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 16
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 17
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 18
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 19
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 20
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 21
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 22
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 23
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 24
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 25
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s58
; VI-NEXT:    v_readlane_b32 s4, v62, 26
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 27
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 28
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 29
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 30
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 31
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 32
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 33
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 34
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 35
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 36
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 37
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 38
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 39
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 40
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 41
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 42
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 43
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 44
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 45
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 46
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 47
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 48
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 49
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 50
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 51
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s60
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 52
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 53
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 54
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 55
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 56
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 57
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_mov_b32_e32 v31, s16
; VI-NEXT:    v_mov_b32_e32 v32, s17
; VI-NEXT:    v_mov_b32_e32 v29, s18
; VI-NEXT:    v_mov_b32_e32 v30, s19
; VI-NEXT:    v_mov_b32_e32 v27, s20
; VI-NEXT:    v_mov_b32_e32 v28, s21
; VI-NEXT:    v_mov_b32_e32 v25, s22
; VI-NEXT:    v_mov_b32_e32 v26, s23
; VI-NEXT:    v_mov_b32_e32 v23, s24
; VI-NEXT:    v_mov_b32_e32 v24, s25
; VI-NEXT:    v_mov_b32_e32 v21, s26
; VI-NEXT:    v_mov_b32_e32 v22, s27
; VI-NEXT:    v_mov_b32_e32 v19, s28
; VI-NEXT:    v_mov_b32_e32 v20, s29
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s45
; VI-NEXT:    v_mov_b32_e32 v15, s42
; VI-NEXT:    v_mov_b32_e32 v16, s43
; VI-NEXT:    v_mov_b32_e32 v13, s40
; VI-NEXT:    v_mov_b32_e32 v14, s41
; VI-NEXT:    v_mov_b32_e32 v11, s14
; VI-NEXT:    v_mov_b32_e32 v12, s15
; VI-NEXT:    v_mov_b32_e32 v9, s12
; VI-NEXT:    v_mov_b32_e32 v10, s13
; VI-NEXT:    v_mov_b32_e32 v7, s10
; VI-NEXT:    v_mov_b32_e32 v8, s11
; VI-NEXT:    v_mov_b32_e32 v5, s8
; VI-NEXT:    v_mov_b32_e32 v6, s9
; VI-NEXT:    v_mov_b32_e32 v3, s6
; VI-NEXT:    v_mov_b32_e32 v4, s7
; VI-NEXT:    v_mov_b32_e32 v2, s5
; VI-NEXT:    v_mov_b32_e32 v61, s71
; VI-NEXT:    v_mov_b32_e32 v33, s69
; VI-NEXT:    v_mov_b32_e32 v60, s70
; VI-NEXT:    v_mov_b32_e32 v50, s68
; VI-NEXT:    v_mov_b32_e32 v59, s67
; VI-NEXT:    v_mov_b32_e32 v58, s66
; VI-NEXT:    v_mov_b32_e32 v48, s64
; VI-NEXT:    v_mov_b32_e32 v52, s65
; VI-NEXT:    v_mov_b32_e32 v56, s55
; VI-NEXT:    v_mov_b32_e32 v57, s54
; VI-NEXT:    v_mov_b32_e32 v46, s53
; VI-NEXT:    v_mov_b32_e32 v47, s51
; VI-NEXT:    v_mov_b32_e32 v45, s52
; VI-NEXT:    v_mov_b32_e32 v43, s50
; VI-NEXT:    v_mov_b32_e32 v44, s87
; VI-NEXT:    v_mov_b32_e32 v49, s86
; VI-NEXT:    v_mov_b32_e32 v42, s84
; VI-NEXT:    v_mov_b32_e32 v51, s85
; VI-NEXT:    v_mov_b32_e32 v38, s83
; VI-NEXT:    v_mov_b32_e32 v55, s82
; VI-NEXT:    v_mov_b32_e32 v36, s81
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, s36
; VI-NEXT:    v_mov_b32_e32 v40, s38
; VI-NEXT:    v_mov_b32_e32 v41, s48
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s62
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s72
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s74
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s76
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s78
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s88
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s90
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v53, s30
; VI-NEXT:    v_mov_b32_e32 v54, s34
; VI-NEXT:  .LBB37_5: ; %end
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v33, v33, v41, s4
; VI-NEXT:    v_perm_b32 v31, v31, v61, s4
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; VI-NEXT:    v_or_b32_e32 v31, v31, v33
; VI-NEXT:    buffer_store_dword v31, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v31, v50, v59, s4
; VI-NEXT:    v_perm_b32 v32, v32, v60, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v31, v32, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v31, v48, v40, s4
; VI-NEXT:    v_perm_b32 v29, v29, v58, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v29, v29, v31
; VI-NEXT:    v_add_u32_e32 v31, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v29, v31, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v29, v30, v52, s4
; VI-NEXT:    v_perm_b32 v30, v56, v57, s4
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    v_add_u32_e32 v30, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v29, v30, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v29, v47, v39, s4
; VI-NEXT:    v_perm_b32 v27, v27, v46, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v27, v27, v29
; VI-NEXT:    v_add_u32_e32 v29, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v27, v29, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v27, v28, v45, s4
; VI-NEXT:    v_perm_b32 v28, v43, v44, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    v_add_u32_e32 v28, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v27, v28, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v27, v42, v54, s4
; VI-NEXT:    v_perm_b32 v25, v25, v49, s4
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v25, v25, v27
; VI-NEXT:    v_add_u32_e32 v27, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v25, v27, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v25, v26, v51, s4
; VI-NEXT:    v_perm_b32 v26, v38, v55, s4
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    v_add_u32_e32 v26, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v25, v26, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v25, v37, v53, s4
; VI-NEXT:    v_perm_b32 v23, v23, v36, s4
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v23, v23, v25
; VI-NEXT:    v_add_u32_e32 v25, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v23, v25, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    v_readlane_b32 s87, v63, 31
; VI-NEXT:    v_readlane_b32 s86, v63, 30
; VI-NEXT:    v_readlane_b32 s85, v63, 29
; VI-NEXT:    v_readlane_b32 s84, v63, 28
; VI-NEXT:    v_readlane_b32 s83, v63, 27
; VI-NEXT:    v_readlane_b32 s82, v63, 26
; VI-NEXT:    v_readlane_b32 s81, v63, 25
; VI-NEXT:    v_readlane_b32 s80, v63, 24
; VI-NEXT:    v_readlane_b32 s71, v63, 23
; VI-NEXT:    v_readlane_b32 s70, v63, 22
; VI-NEXT:    v_readlane_b32 s69, v63, 21
; VI-NEXT:    v_readlane_b32 s68, v63, 20
; VI-NEXT:    v_readlane_b32 s67, v63, 19
; VI-NEXT:    v_readlane_b32 s66, v63, 18
; VI-NEXT:    v_readlane_b32 s65, v63, 17
; VI-NEXT:    v_readlane_b32 s64, v63, 16
; VI-NEXT:    v_readlane_b32 s55, v63, 15
; VI-NEXT:    v_readlane_b32 s54, v63, 14
; VI-NEXT:    v_readlane_b32 s53, v63, 13
; VI-NEXT:    v_readlane_b32 s52, v63, 12
; VI-NEXT:    v_readlane_b32 s51, v63, 11
; VI-NEXT:    v_readlane_b32 s50, v63, 10
; VI-NEXT:    v_readlane_b32 s49, v63, 9
; VI-NEXT:    v_readlane_b32 s48, v63, 8
; VI-NEXT:    v_readlane_b32 s39, v63, 7
; VI-NEXT:    v_readlane_b32 s38, v63, 6
; VI-NEXT:    v_readlane_b32 s37, v63, 5
; VI-NEXT:    v_readlane_b32 s36, v63, 4
; VI-NEXT:    v_readlane_b32 s35, v63, 3
; VI-NEXT:    v_readlane_b32 s34, v63, 2
; VI-NEXT:    v_readlane_b32 s31, v63, 1
; VI-NEXT:    v_readlane_b32 s30, v63, 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s4
; VI-NEXT:    v_perm_b32 v24, v34, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    v_add_u32_e32 v24, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v23, v24, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s4
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v24, s4
; VI-NEXT:    v_or_b32_e32 v21, v21, v23
; VI-NEXT:    v_add_u32_e32 v23, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v21, v23, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    v_add_u32_e32 v22, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v21, v22, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v19, v22, s4
; VI-NEXT:    v_or_b32_e32 v19, v19, v21
; VI-NEXT:    v_add_u32_e32 v21, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v19, v21, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v20, v21, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    v_add_u32_e32 v20, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v19, v20, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v17, v20, s4
; VI-NEXT:    v_or_b32_e32 v17, v17, v19
; VI-NEXT:    v_add_u32_e32 v19, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v17, v19, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s4
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v18, v19, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    v_add_u32_e32 v18, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v17, v18, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s4
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v15, v18, s4
; VI-NEXT:    v_or_b32_e32 v15, v15, v17
; VI-NEXT:    v_add_u32_e32 v17, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v15, v17, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v16, v15, s4
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v16, v17, s4
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v15, v16, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v16, v15, s4
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v13, v16, s4
; VI-NEXT:    v_or_b32_e32 v13, v13, v15
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v13, v15, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v14, v13, s4
; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v14, v14, v15, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v13, v14, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v11, v13, s4
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v14, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_e32 v11, v11, v13
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v11, v13, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v12, v11, s4
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v12, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v11, v12, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v9, v11, s4
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v12, v11, s4
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_or_b32_e32 v9, v9, v11
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v9, v11, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v10, v9, s4
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v10, v10, v11, s4
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v9, v10, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v7, v9, s4
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v10, v9, s4
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_e32 v7, v7, v9
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v7, v9, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s4
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v9, s4
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v7, v8, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v5, v7, s4
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s4
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v5, v5, v7
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v5, v7, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s4
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v6, v7, s4
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v3, v5, s4
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v3, v3, v5
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v3, v5, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v4, v5, s4
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v3, s4
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v1, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v128i8_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v63, s30, 0
; GFX9-NEXT:    v_writelane_b32 v63, s31, 1
; GFX9-NEXT:    v_writelane_b32 v63, s34, 2
; GFX9-NEXT:    v_writelane_b32 v63, s35, 3
; GFX9-NEXT:    v_writelane_b32 v63, s36, 4
; GFX9-NEXT:    v_writelane_b32 v63, s37, 5
; GFX9-NEXT:    v_writelane_b32 v63, s38, 6
; GFX9-NEXT:    v_writelane_b32 v63, s39, 7
; GFX9-NEXT:    v_writelane_b32 v63, s48, 8
; GFX9-NEXT:    v_writelane_b32 v63, s49, 9
; GFX9-NEXT:    v_writelane_b32 v63, s50, 10
; GFX9-NEXT:    v_writelane_b32 v63, s51, 11
; GFX9-NEXT:    v_writelane_b32 v63, s52, 12
; GFX9-NEXT:    v_writelane_b32 v63, s53, 13
; GFX9-NEXT:    v_writelane_b32 v63, s54, 14
; GFX9-NEXT:    v_writelane_b32 v63, s55, 15
; GFX9-NEXT:    v_writelane_b32 v63, s64, 16
; GFX9-NEXT:    v_writelane_b32 v63, s65, 17
; GFX9-NEXT:    v_writelane_b32 v63, s66, 18
; GFX9-NEXT:    v_writelane_b32 v63, s67, 19
; GFX9-NEXT:    v_writelane_b32 v63, s68, 20
; GFX9-NEXT:    v_writelane_b32 v63, s69, 21
; GFX9-NEXT:    v_writelane_b32 v63, s70, 22
; GFX9-NEXT:    v_writelane_b32 v63, s71, 23
; GFX9-NEXT:    v_writelane_b32 v63, s80, 24
; GFX9-NEXT:    v_writelane_b32 v63, s81, 25
; GFX9-NEXT:    v_writelane_b32 v63, s82, 26
; GFX9-NEXT:    v_writelane_b32 v63, s83, 27
; GFX9-NEXT:    v_writelane_b32 v63, s84, 28
; GFX9-NEXT:    v_writelane_b32 v63, s85, 29
; GFX9-NEXT:    v_writelane_b32 v63, s86, 30
; GFX9-NEXT:    v_writelane_b32 v63, s87, 31
; GFX9-NEXT:    v_writelane_b32 v63, s96, 32
; GFX9-NEXT:    v_writelane_b32 v63, s97, 33
; GFX9-NEXT:    v_writelane_b32 v63, s98, 34
; GFX9-NEXT:    v_readfirstlane_b32 s44, v19
; GFX9-NEXT:    v_writelane_b32 v63, s99, 35
; GFX9-NEXT:    v_readfirstlane_b32 s5, v18
; GFX9-NEXT:    v_readfirstlane_b32 s4, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s6, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s8, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s10, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s12, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s14, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s40, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s42, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    s_cmp_lg_u32 s44, 0
; GFX9-NEXT:    v_readfirstlane_b32 s44, v1
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; GFX9-NEXT:    s_cbranch_scc0 .LBB37_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 49
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s6, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s6, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s8, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s8, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s11, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s11, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s11, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s10, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s10, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s13, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s13, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s13, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s12, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s12, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s15, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s15, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s15, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s14, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s14, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s41, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s41, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s41, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s40, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s40, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s43, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s43, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s43, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s42, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s42, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s45, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s45, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s45, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s44, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s44, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s29, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s28, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 0
; GFX9-NEXT:    s_lshr_b32 s82, s27, 24
; GFX9-NEXT:    s_lshr_b32 s83, s27, 16
; GFX9-NEXT:    s_lshr_b32 s85, s27, 8
; GFX9-NEXT:    s_lshr_b32 s84, s26, 16
; GFX9-NEXT:    s_lshr_b32 s86, s26, 8
; GFX9-NEXT:    s_lshr_b32 s87, s25, 24
; GFX9-NEXT:    s_lshr_b32 s96, s25, 16
; GFX9-NEXT:    s_lshr_b32 s98, s25, 8
; GFX9-NEXT:    s_lshr_b32 s97, s24, 16
; GFX9-NEXT:    s_lshr_b32 s99, s24, 8
; GFX9-NEXT:    s_lshr_b32 s38, s23, 24
; GFX9-NEXT:    s_lshr_b32 s39, s23, 16
; GFX9-NEXT:    s_lshr_b32 s49, s23, 8
; GFX9-NEXT:    s_lshr_b32 s48, s22, 16
; GFX9-NEXT:    s_lshr_b32 s50, s22, 8
; GFX9-NEXT:    s_lshr_b32 s51, s21, 24
; GFX9-NEXT:    s_lshr_b32 s52, s21, 16
; GFX9-NEXT:    s_lshr_b32 s54, s21, 8
; GFX9-NEXT:    s_lshr_b32 s53, s20, 16
; GFX9-NEXT:    s_lshr_b32 s55, s20, 8
; GFX9-NEXT:    s_lshr_b32 s64, s19, 24
; GFX9-NEXT:    s_lshr_b32 s65, s19, 16
; GFX9-NEXT:    s_lshr_b32 s67, s19, 8
; GFX9-NEXT:    s_lshr_b32 s66, s18, 16
; GFX9-NEXT:    s_lshr_b32 s68, s18, 8
; GFX9-NEXT:    s_lshr_b32 s69, s17, 24
; GFX9-NEXT:    s_lshr_b32 s70, s17, 16
; GFX9-NEXT:    s_lshr_b32 s80, s17, 8
; GFX9-NEXT:    s_lshr_b32 s71, s16, 16
; GFX9-NEXT:    s_lshr_b32 s81, s16, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[16:17], 24
; GFX9-NEXT:    s_cbranch_execnz .LBB37_4
; GFX9-NEXT:  .LBB37_2: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e64 v2, s5, 1.0
; GFX9-NEXT:    v_add_f32_e64 v1, s4, 1.0
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[1:2]
; GFX9-NEXT:    v_add_f32_e64 v4, s7, 1.0
; GFX9-NEXT:    v_add_f32_e64 v3, s6, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[3:4]
; GFX9-NEXT:    v_add_f32_e64 v6, s9, 1.0
; GFX9-NEXT:    v_add_f32_e64 v5, s8, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
; GFX9-NEXT:    v_add_f32_e64 v8, s11, 1.0
; GFX9-NEXT:    v_add_f32_e64 v7, s10, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
; GFX9-NEXT:    v_add_f32_e64 v10, s13, 1.0
; GFX9-NEXT:    v_add_f32_e64 v9, s12, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[9:10]
; GFX9-NEXT:    v_add_f32_e64 v12, s15, 1.0
; GFX9-NEXT:    v_add_f32_e64 v11, s14, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[11:12]
; GFX9-NEXT:    v_add_f32_e64 v14, s41, 1.0
; GFX9-NEXT:    v_add_f32_e64 v13, s40, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[13:14]
; GFX9-NEXT:    v_add_f32_e64 v22, s43, 1.0
; GFX9-NEXT:    v_add_f32_e64 v21, s42, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[21:22]
; GFX9-NEXT:    v_add_f32_e64 v24, s45, 1.0
; GFX9-NEXT:    v_add_f32_e64 v23, s44, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[23:24]
; GFX9-NEXT:    v_add_f32_e64 v26, s29, 1.0
; GFX9-NEXT:    v_add_f32_e64 v25, s28, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v3
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v3
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v5
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v5
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v7
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v7
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v9
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v9
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v22
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v22
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v22
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v21
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v21
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v26
; GFX9-NEXT:    v_add_f32_e64 v28, s27, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v26
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v28
; GFX9-NEXT:    v_add_f32_e64 v32, s23, 1.0
; GFX9-NEXT:    v_add_f32_e64 v31, s22, 1.0
; GFX9-NEXT:    v_add_f32_e64 v30, s25, 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v28
; GFX9-NEXT:    v_add_f32_e64 v34, s21, 1.0
; GFX9-NEXT:    v_add_f32_e64 v33, s20, 1.0
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v30
; GFX9-NEXT:    v_add_f32_e64 v49, s17, 1.0
; GFX9-NEXT:    v_add_f32_e64 v36, s19, 1.0
; GFX9-NEXT:    v_add_f32_e64 v35, s18, 1.0
; GFX9-NEXT:    v_add_f32_e64 v27, s26, 1.0
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[33:34]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v30
; GFX9-NEXT:    v_add_f32_e64 v48, s16, 1.0
; GFX9-NEXT:    v_add_f32_e64 v29, s24, 1.0
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[27:28]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[35:36]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 24, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 24, v36
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 24, v49
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[29:30]
; GFX9-NEXT:    v_lshrrev_b64 v[42:43], 24, v[48:49]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v32
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 16, v34
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v36
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 16, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 8, v23
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v26
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 16, v25
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v25
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v28
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 8, v30
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 16, v29
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 8, v29
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 16, v31
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v31
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 8, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 16, v33
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 8, v33
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 8, v36
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 16, v35
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 8, v35
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 8, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 16, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 8, v48
; GFX9-NEXT:    s_branch .LBB37_5
; GFX9-NEXT:  .LBB37_3:
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr81
; GFX9-NEXT:    ; implicit-def: $sgpr71
; GFX9-NEXT:    ; implicit-def: $sgpr80
; GFX9-NEXT:    ; implicit-def: $sgpr70
; GFX9-NEXT:    ; implicit-def: $sgpr69
; GFX9-NEXT:    ; implicit-def: $sgpr68
; GFX9-NEXT:    ; implicit-def: $sgpr66
; GFX9-NEXT:    ; implicit-def: $sgpr67
; GFX9-NEXT:    ; implicit-def: $sgpr65
; GFX9-NEXT:    ; implicit-def: $sgpr64
; GFX9-NEXT:    ; implicit-def: $sgpr55
; GFX9-NEXT:    ; implicit-def: $sgpr53
; GFX9-NEXT:    ; implicit-def: $sgpr54
; GFX9-NEXT:    ; implicit-def: $sgpr52
; GFX9-NEXT:    ; implicit-def: $sgpr51
; GFX9-NEXT:    ; implicit-def: $sgpr50
; GFX9-NEXT:    ; implicit-def: $sgpr48
; GFX9-NEXT:    ; implicit-def: $sgpr49
; GFX9-NEXT:    ; implicit-def: $sgpr39
; GFX9-NEXT:    ; implicit-def: $sgpr38
; GFX9-NEXT:    ; implicit-def: $sgpr99
; GFX9-NEXT:    ; implicit-def: $sgpr97
; GFX9-NEXT:    ; implicit-def: $sgpr98
; GFX9-NEXT:    ; implicit-def: $sgpr96
; GFX9-NEXT:    ; implicit-def: $sgpr87
; GFX9-NEXT:    ; implicit-def: $sgpr86
; GFX9-NEXT:    ; implicit-def: $sgpr84
; GFX9-NEXT:    ; implicit-def: $sgpr85
; GFX9-NEXT:    ; implicit-def: $sgpr83
; GFX9-NEXT:    ; implicit-def: $sgpr82
; GFX9-NEXT:    ; implicit-def: $sgpr36
; GFX9-NEXT:    ; implicit-def: $sgpr34
; GFX9-NEXT:    ; implicit-def: $sgpr30
; GFX9-NEXT:    ; implicit-def: $sgpr94
; GFX9-NEXT:    ; implicit-def: $sgpr92
; GFX9-NEXT:    ; implicit-def: $sgpr90
; GFX9-NEXT:    ; implicit-def: $sgpr88
; GFX9-NEXT:    ; implicit-def: $sgpr78
; GFX9-NEXT:    ; implicit-def: $sgpr76
; GFX9-NEXT:    ; implicit-def: $sgpr74
; GFX9-NEXT:    ; implicit-def: $sgpr72
; GFX9-NEXT:    ; implicit-def: $sgpr62
; GFX9-NEXT:    ; implicit-def: $sgpr60
; GFX9-NEXT:    ; implicit-def: $sgpr58
; GFX9-NEXT:    ; implicit-def: $sgpr56
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    s_branch .LBB37_2
; GFX9-NEXT:  .LBB37_4:
; GFX9-NEXT:    v_mov_b32_e32 v53, s46
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s56
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_mov_b32_e32 v39, s96
; GFX9-NEXT:    v_readlane_b32 s4, v62, 0
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s87
; GFX9-NEXT:    v_mov_b32_e32 v46, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 1
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s83
; GFX9-NEXT:    v_mov_b32_e32 v47, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 2
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s82
; GFX9-NEXT:    v_mov_b32_e32 v45, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 3
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 4
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 5
; GFX9-NEXT:    v_mov_b32_e32 v43, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 6
; GFX9-NEXT:    v_mov_b32_e32 v44, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 7
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 8
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 9
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 10
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 11
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 12
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 13
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 14
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 15
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 16
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 17
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 18
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 19
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 20
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 21
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 22
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 23
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 24
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s58
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 25
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 26
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 27
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 28
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 29
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 30
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 31
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 32
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 33
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 34
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 35
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 36
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 37
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 38
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 39
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 40
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 41
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 42
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 43
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 44
; GFX9-NEXT:    v_mov_b32_e32 v15, s70
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 45
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s69
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 46
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s65
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 47
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s64
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s60
; GFX9-NEXT:    v_readlane_b32 s4, v62, 48
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s52
; GFX9-NEXT:    v_mov_b32_e32 v37, s39
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 49
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v37, s38
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_mov_b32_e32 v48, s16
; GFX9-NEXT:    v_mov_b32_e32 v49, s17
; GFX9-NEXT:    v_mov_b32_e32 v35, s18
; GFX9-NEXT:    v_mov_b32_e32 v36, s19
; GFX9-NEXT:    v_mov_b32_e32 v33, s20
; GFX9-NEXT:    v_mov_b32_e32 v34, s21
; GFX9-NEXT:    v_mov_b32_e32 v31, s22
; GFX9-NEXT:    v_mov_b32_e32 v32, s23
; GFX9-NEXT:    v_mov_b32_e32 v29, s24
; GFX9-NEXT:    v_mov_b32_e32 v30, s25
; GFX9-NEXT:    v_mov_b32_e32 v27, s26
; GFX9-NEXT:    v_mov_b32_e32 v28, s27
; GFX9-NEXT:    v_mov_b32_e32 v25, s28
; GFX9-NEXT:    v_mov_b32_e32 v26, s29
; GFX9-NEXT:    v_mov_b32_e32 v23, s44
; GFX9-NEXT:    v_mov_b32_e32 v24, s45
; GFX9-NEXT:    v_mov_b32_e32 v21, s42
; GFX9-NEXT:    v_mov_b32_e32 v22, s43
; GFX9-NEXT:    v_mov_b32_e32 v13, s40
; GFX9-NEXT:    v_mov_b32_e32 v14, s41
; GFX9-NEXT:    v_mov_b32_e32 v11, s14
; GFX9-NEXT:    v_mov_b32_e32 v12, s15
; GFX9-NEXT:    v_mov_b32_e32 v9, s12
; GFX9-NEXT:    v_mov_b32_e32 v10, s13
; GFX9-NEXT:    v_mov_b32_e32 v7, s10
; GFX9-NEXT:    v_mov_b32_e32 v8, s11
; GFX9-NEXT:    v_mov_b32_e32 v5, s8
; GFX9-NEXT:    v_mov_b32_e32 v6, s9
; GFX9-NEXT:    v_mov_b32_e32 v3, s6
; GFX9-NEXT:    v_mov_b32_e32 v4, s7
; GFX9-NEXT:    v_mov_b32_e32 v2, s5
; GFX9-NEXT:    v_mov_b32_e32 v20, s81
; GFX9-NEXT:    v_mov_b32_e32 v55, s71
; GFX9-NEXT:    v_mov_b32_e32 v19, s80
; GFX9-NEXT:    v_mov_b32_e32 v18, s68
; GFX9-NEXT:    v_mov_b32_e32 v52, s66
; GFX9-NEXT:    v_mov_b32_e32 v17, s67
; GFX9-NEXT:    v_mov_b32_e32 v61, s55
; GFX9-NEXT:    v_mov_b32_e32 v51, s53
; GFX9-NEXT:    v_mov_b32_e32 v16, s54
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v60, s50
; GFX9-NEXT:    v_mov_b32_e32 v50, s48
; GFX9-NEXT:    v_mov_b32_e32 v15, s49
; GFX9-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v37, s99
; GFX9-NEXT:    v_mov_b32_e32 v59, s97
; GFX9-NEXT:    v_mov_b32_e32 v38, s98
; GFX9-NEXT:    v_mov_b32_e32 v57, s86
; GFX9-NEXT:    v_mov_b32_e32 v58, s84
; GFX9-NEXT:    v_mov_b32_e32 v56, s85
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v39, s94
; GFX9-NEXT:    v_mov_b32_e32 v40, s30
; GFX9-NEXT:    v_mov_b32_e32 v41, s34
; GFX9-NEXT:    v_mov_b32_e32 v42, s36
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s62
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s72
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s74
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s76
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s78
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s88
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v53, s90
; GFX9-NEXT:    v_mov_b32_e32 v54, s92
; GFX9-NEXT:  .LBB37_5: ; %end
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v32, v32, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v19, v49, v19, s4
; GFX9-NEXT:    v_perm_b32 v16, v34, v16, s4
; GFX9-NEXT:    v_perm_b32 v34, v50, v39, s4
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v30, v30, v38, s4
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v18, v35, v18, s4
; GFX9-NEXT:    v_perm_b32 v35, v51, v40, s4
; GFX9-NEXT:    v_perm_b32 v55, v55, v42, s4
; GFX9-NEXT:    v_perm_b32 v20, v48, v20, s4
; GFX9-NEXT:    v_perm_b32 v48, v52, v41, s4
; GFX9-NEXT:    v_perm_b32 v17, v36, v17, s4
; GFX9-NEXT:    v_perm_b32 v33, v33, v61, s4
; GFX9-NEXT:    v_perm_b32 v31, v31, v60, s4
; GFX9-NEXT:    v_perm_b32 v36, v59, v54, s4
; GFX9-NEXT:    v_perm_b32 v29, v29, v37, s4
; GFX9-NEXT:    v_perm_b32 v37, v58, v53, s4
; GFX9-NEXT:    v_perm_b32 v27, v27, v57, s4
; GFX9-NEXT:    v_perm_b32 v28, v28, v56, s4
; GFX9-NEXT:    v_perm_b32 v25, v25, v46, s4
; GFX9-NEXT:    v_perm_b32 v26, v26, v45, s4
; GFX9-NEXT:    v_perm_b32 v23, v23, v43, s4
; GFX9-NEXT:    v_readlane_b32 s99, v63, 35
; GFX9-NEXT:    v_readlane_b32 s98, v63, 34
; GFX9-NEXT:    v_readlane_b32 s97, v63, 33
; GFX9-NEXT:    v_readlane_b32 s96, v63, 32
; GFX9-NEXT:    v_readlane_b32 s87, v63, 31
; GFX9-NEXT:    v_readlane_b32 s86, v63, 30
; GFX9-NEXT:    v_readlane_b32 s85, v63, 29
; GFX9-NEXT:    v_readlane_b32 s84, v63, 28
; GFX9-NEXT:    v_readlane_b32 s83, v63, 27
; GFX9-NEXT:    v_readlane_b32 s82, v63, 26
; GFX9-NEXT:    v_readlane_b32 s81, v63, 25
; GFX9-NEXT:    v_readlane_b32 s80, v63, 24
; GFX9-NEXT:    v_readlane_b32 s71, v63, 23
; GFX9-NEXT:    v_readlane_b32 s70, v63, 22
; GFX9-NEXT:    v_readlane_b32 s69, v63, 21
; GFX9-NEXT:    v_readlane_b32 s68, v63, 20
; GFX9-NEXT:    v_readlane_b32 s67, v63, 19
; GFX9-NEXT:    v_readlane_b32 s66, v63, 18
; GFX9-NEXT:    v_readlane_b32 s65, v63, 17
; GFX9-NEXT:    v_readlane_b32 s64, v63, 16
; GFX9-NEXT:    v_readlane_b32 s55, v63, 15
; GFX9-NEXT:    v_readlane_b32 s54, v63, 14
; GFX9-NEXT:    v_readlane_b32 s53, v63, 13
; GFX9-NEXT:    v_readlane_b32 s52, v63, 12
; GFX9-NEXT:    v_readlane_b32 s51, v63, 11
; GFX9-NEXT:    v_readlane_b32 s50, v63, 10
; GFX9-NEXT:    v_readlane_b32 s49, v63, 9
; GFX9-NEXT:    v_readlane_b32 s48, v63, 8
; GFX9-NEXT:    v_readlane_b32 s39, v63, 7
; GFX9-NEXT:    v_readlane_b32 s38, v63, 6
; GFX9-NEXT:    v_readlane_b32 s37, v63, 5
; GFX9-NEXT:    v_readlane_b32 s36, v63, 4
; GFX9-NEXT:    v_readlane_b32 s35, v63, 3
; GFX9-NEXT:    v_readlane_b32 s34, v63, 2
; GFX9-NEXT:    v_readlane_b32 s31, v63, 1
; GFX9-NEXT:    v_readlane_b32 s30, v63, 0
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v24, v24, v15, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v39, v44, v49, s4
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v38, v47, v38, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v49, v15, v49, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v21, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v22, v15, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v15, v15, v50, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v50, 16, v55
; GFX9-NEXT:    v_or_b32_e32 v20, v20, v50
; GFX9-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v20, v50, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
; GFX9-NEXT:    buffer_store_dword v19, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshlrev_b32_e32 v19, 16, v48
; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
; GFX9-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v19, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX9-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v35
; GFX9-NEXT:    v_or_b32_e32 v17, v33, v17
; GFX9-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v17, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v34
; GFX9-NEXT:    v_or_b32_e32 v16, v31, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v32, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v36
; GFX9-NEXT:    v_or_b32_e32 v16, v29, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v30, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v37
; GFX9-NEXT:    v_or_b32_e32 v16, v27, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v28, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v38
; GFX9-NEXT:    v_or_b32_e32 v16, v25, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v26, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v39
; GFX9-NEXT:    v_or_b32_e32 v16, v23, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v24, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v49
; GFX9-NEXT:    v_or_b32_e32 v16, v21, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v22, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v16, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v15
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v14, v15, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v11, v13, s4
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v13
; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s4
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v12, v12, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v9, v11, s4
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v11
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v10, v11, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v7, v9, s4
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v9
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s4
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v9, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v7, s4
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s4
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v7, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v3, v5, s4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v5
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v4, v5, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v3, s4
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v128i8_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:72
; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:76
; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v74, s30, 0
; GFX11-NEXT:    v_writelane_b32 v75, s96, 0
; GFX11-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_writelane_b32 v74, s31, 1
; GFX11-NEXT:    v_writelane_b32 v75, s97, 1
; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
; GFX11-NEXT:    v_writelane_b32 v74, s34, 2
; GFX11-NEXT:    v_writelane_b32 v75, s98, 2
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_writelane_b32 v74, s35, 3
; GFX11-NEXT:    v_writelane_b32 v75, s99, 3
; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
; GFX11-NEXT:    v_writelane_b32 v74, s36, 4
; GFX11-NEXT:    v_writelane_b32 v75, s100, 4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_writelane_b32 v74, s37, 5
; GFX11-NEXT:    v_writelane_b32 v75, s101, 5
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:4
; GFX11-NEXT:    scratch_store_b32 off, v73, s32
; GFX11-NEXT:    v_writelane_b32 v74, s38, 6
; GFX11-NEXT:    v_writelane_b32 v75, s102, 6
; GFX11-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
; GFX11-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
; GFX11-NEXT:    v_writelane_b32 v74, s39, 7
; GFX11-NEXT:    v_writelane_b32 v75, s103, 7
; GFX11-NEXT:    v_writelane_b32 v74, s48, 8
; GFX11-NEXT:    v_writelane_b32 v75, s104, 8
; GFX11-NEXT:    v_writelane_b32 v74, s49, 9
; GFX11-NEXT:    v_writelane_b32 v74, s50, 10
; GFX11-NEXT:    v_writelane_b32 v74, s51, 11
; GFX11-NEXT:    v_writelane_b32 v74, s52, 12
; GFX11-NEXT:    v_writelane_b32 v74, s53, 13
; GFX11-NEXT:    v_writelane_b32 v74, s54, 14
; GFX11-NEXT:    v_writelane_b32 v74, s55, 15
; GFX11-NEXT:    v_writelane_b32 v74, s64, 16
; GFX11-NEXT:    v_writelane_b32 v74, s65, 17
; GFX11-NEXT:    v_writelane_b32 v74, s66, 18
; GFX11-NEXT:    v_writelane_b32 v74, s67, 19
; GFX11-NEXT:    v_writelane_b32 v74, s68, 20
; GFX11-NEXT:    v_writelane_b32 v74, s69, 21
; GFX11-NEXT:    v_writelane_b32 v74, s70, 22
; GFX11-NEXT:    v_writelane_b32 v74, s71, 23
; GFX11-NEXT:    v_writelane_b32 v74, s80, 24
; GFX11-NEXT:    v_writelane_b32 v74, s81, 25
; GFX11-NEXT:    v_writelane_b32 v74, s82, 26
; GFX11-NEXT:    v_writelane_b32 v74, s83, 27
; GFX11-NEXT:    v_writelane_b32 v74, s84, 28
; GFX11-NEXT:    v_writelane_b32 v74, s85, 29
; GFX11-NEXT:    v_writelane_b32 v74, s86, 30
; GFX11-NEXT:    v_writelane_b32 v74, s87, 31
; GFX11-NEXT:    s_cbranch_scc0 .LBB37_3
; GFX11-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s27, 24
; GFX11-NEXT:    v_writelane_b32 v77, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
; GFX11-NEXT:    s_lshr_b32 s34, s27, 16
; GFX11-NEXT:    s_lshr_b32 s36, s27, 8
; GFX11-NEXT:    s_lshr_b32 s35, s26, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
; GFX11-NEXT:    s_lshr_b32 s37, s26, 8
; GFX11-NEXT:    s_lshr_b32 s38, s25, 24
; GFX11-NEXT:    s_lshr_b32 s39, s25, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s4, 16
; GFX11-NEXT:    s_lshr_b32 s49, s25, 8
; GFX11-NEXT:    s_lshr_b32 s48, s24, 16
; GFX11-NEXT:    s_lshr_b32 s50, s24, 8
; GFX11-NEXT:    v_writelane_b32 v77, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s4, 8
; GFX11-NEXT:    s_lshr_b32 s51, s23, 24
; GFX11-NEXT:    s_lshr_b32 s52, s23, 16
; GFX11-NEXT:    s_lshr_b32 s54, s23, 8
; GFX11-NEXT:    v_writelane_b32 v77, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
; GFX11-NEXT:    s_lshr_b32 s53, s22, 16
; GFX11-NEXT:    s_lshr_b32 s55, s22, 8
; GFX11-NEXT:    s_lshr_b32 s64, s21, 24
; GFX11-NEXT:    v_writelane_b32 v77, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
; GFX11-NEXT:    s_lshr_b32 s65, s21, 16
; GFX11-NEXT:    s_lshr_b32 s67, s21, 8
; GFX11-NEXT:    s_lshr_b32 s66, s20, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
; GFX11-NEXT:    s_lshr_b32 s68, s20, 8
; GFX11-NEXT:    s_lshr_b32 s69, s19, 24
; GFX11-NEXT:    s_lshr_b32 s70, s19, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s6, 16
; GFX11-NEXT:    s_lshr_b32 s80, s19, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s6, 8
; GFX11-NEXT:    s_lshr_b32 s71, s18, 16
; GFX11-NEXT:    s_lshr_b32 s81, s18, 8
; GFX11-NEXT:    s_lshr_b32 s82, s17, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
; GFX11-NEXT:    s_lshr_b32 s83, s17, 16
; GFX11-NEXT:    s_lshr_b32 s85, s17, 8
; GFX11-NEXT:    s_lshr_b32 s84, s16, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
; GFX11-NEXT:    s_lshr_b32 s86, s16, 8
; GFX11-NEXT:    s_lshr_b32 s87, s3, 24
; GFX11-NEXT:    s_lshr_b32 s96, s3, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
; GFX11-NEXT:    s_lshr_b32 s98, s3, 8
; GFX11-NEXT:    s_lshr_b32 s97, s2, 16
; GFX11-NEXT:    s_lshr_b32 s99, s2, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s8, 16
; GFX11-NEXT:    s_lshr_b32 s100, s1, 24
; GFX11-NEXT:    s_lshr_b32 s101, s1, 16
; GFX11-NEXT:    s_lshr_b32 s103, s1, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s8, 8
; GFX11-NEXT:    s_lshr_b32 s102, s0, 16
; GFX11-NEXT:    s_lshr_b32 s104, s0, 8
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[4:5], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s11, 24
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[6:7], 24
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[8:9], 24
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[10:11], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[12:13], 24
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[14:15], 24
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[40:41], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s11, 8
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s10, 16
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s10, 8
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s13, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s12, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s12, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s15, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s15, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s15, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s14, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s14, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s41, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s41, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s41, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s40, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s40, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 0
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-NEXT:    s_cbranch_vccnz .LBB37_4
; GFX11-NEXT:  .LBB37_2: ; %cmp.true
; GFX11-NEXT:    v_add_f32_e64 v12, s15, 1.0
; GFX11-NEXT:    v_add_f32_e64 v11, s14, 1.0
; GFX11-NEXT:    v_add_f32_e64 v14, s41, 1.0
; GFX11-NEXT:    v_add_f32_e64 v13, s40, 1.0
; GFX11-NEXT:    v_add_f32_e64 v16, s29, 1.0
; GFX11-NEXT:    v_add_f32_e64 v15, s28, 1.0
; GFX11-NEXT:    v_add_f32_e64 v26, s19, 1.0
; GFX11-NEXT:    v_add_f32_e64 v25, s18, 1.0
; GFX11-NEXT:    v_add_f32_e64 v18, s27, 1.0
; GFX11-NEXT:    v_add_f32_e64 v17, s26, 1.0
; GFX11-NEXT:    v_add_f32_e64 v4, s7, 1.0
; GFX11-NEXT:    v_add_f32_e64 v3, s6, 1.0
; GFX11-NEXT:    v_add_f32_e64 v28, s17, 1.0
; GFX11-NEXT:    v_add_f32_e64 v27, s16, 1.0
; GFX11-NEXT:    v_add_f32_e64 v20, s25, 1.0
; GFX11-NEXT:    v_add_f32_e64 v19, s24, 1.0
; GFX11-NEXT:    v_add_f32_e64 v6, s9, 1.0
; GFX11-NEXT:    v_add_f32_e64 v5, s8, 1.0
; GFX11-NEXT:    v_add_f32_e64 v32, s3, 1.0
; GFX11-NEXT:    v_add_f32_e64 v31, s2, 1.0
; GFX11-NEXT:    v_add_f32_e64 v22, s23, 1.0
; GFX11-NEXT:    v_add_f32_e64 v21, s22, 1.0
; GFX11-NEXT:    v_add_f32_e64 v8, s11, 1.0
; GFX11-NEXT:    v_add_f32_e64 v7, s10, 1.0
; GFX11-NEXT:    v_lshrrev_b64 v[48:49], 24, v[11:12]
; GFX11-NEXT:    v_add_f32_e64 v34, s1, 1.0
; GFX11-NEXT:    v_add_f32_e64 v33, s0, 1.0
; GFX11-NEXT:    v_add_f32_e64 v24, s21, 1.0
; GFX11-NEXT:    v_add_f32_e64 v23, s20, 1.0
; GFX11-NEXT:    v_add_f32_e64 v10, s13, 1.0
; GFX11-NEXT:    v_add_f32_e64 v2, s5, 1.0
; GFX11-NEXT:    v_add_f32_e64 v1, s4, 1.0
; GFX11-NEXT:    v_add_f32_e64 v9, s12, 1.0
; GFX11-NEXT:    v_lshrrev_b64 v[49:50], 24, v[13:14]
; GFX11-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-NEXT:    v_lshrrev_b64 v[35:36], 24, v[3:4]
; GFX11-NEXT:    v_lshrrev_b64 v[51:52], 24, v[17:18]
; GFX11-NEXT:    v_lshrrev_b64 v[64:65], 24, v[25:26]
; GFX11-NEXT:    v_lshrrev_b64 v[36:37], 24, v[5:6]
; GFX11-NEXT:    v_lshrrev_b64 v[52:53], 24, v[19:20]
; GFX11-NEXT:    v_lshrrev_b64 v[65:66], 24, v[27:28]
; GFX11-NEXT:    v_lshrrev_b64 v[37:38], 24, v[7:8]
; GFX11-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX11-NEXT:    v_lshrrev_b64 v[66:67], 24, v[31:32]
; GFX11-NEXT:    v_lshrrev_b64 v[29:30], 24, v[1:2]
; GFX11-NEXT:    v_lshrrev_b64 v[38:39], 24, v[9:10]
; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[33:34]
; GFX11-NEXT:    v_lshrrev_b32_e32 v70, 24, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v69, 16, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 8, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 8, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 24, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 16, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 8, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 16, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 8, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 24, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 8, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 16, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 8, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 24, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 16, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 8, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 8, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 24, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 16, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 8, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 16, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 8, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 24, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 16, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 8, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 8, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 24, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 16, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 16, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 8, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 24, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 16, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 8, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 16, v15
; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 8, v15
; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 24, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 16, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 8, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 16, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 24, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 16, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 8, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 16, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 8, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 24, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 16, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 8, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 16, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 24, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 16, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 8, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 8, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 24, v26
; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 16, v26
; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 8, v26
; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 16, v25
; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 8, v25
; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 24, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 16, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 8, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 16, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 8, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 24, v32
; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 16, v32
; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 8, v32
; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 16, v31
; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 8, v31
; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 24, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 16, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 8, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 16, v33
; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 8, v33
; GFX11-NEXT:    s_branch .LBB37_5
; GFX11-NEXT:  .LBB37_3:
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr104
; GFX11-NEXT:    ; implicit-def: $sgpr102
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr103
; GFX11-NEXT:    ; implicit-def: $sgpr101
; GFX11-NEXT:    ; implicit-def: $sgpr100
; GFX11-NEXT:    ; implicit-def: $sgpr99
; GFX11-NEXT:    ; implicit-def: $sgpr97
; GFX11-NEXT:    ; implicit-def: $sgpr44
; GFX11-NEXT:    ; implicit-def: $sgpr98
; GFX11-NEXT:    ; implicit-def: $sgpr96
; GFX11-NEXT:    ; implicit-def: $sgpr87
; GFX11-NEXT:    ; implicit-def: $sgpr86
; GFX11-NEXT:    ; implicit-def: $sgpr84
; GFX11-NEXT:    ; implicit-def: $sgpr46
; GFX11-NEXT:    ; implicit-def: $sgpr85
; GFX11-NEXT:    ; implicit-def: $sgpr83
; GFX11-NEXT:    ; implicit-def: $sgpr82
; GFX11-NEXT:    ; implicit-def: $sgpr81
; GFX11-NEXT:    ; implicit-def: $sgpr71
; GFX11-NEXT:    ; implicit-def: $sgpr56
; GFX11-NEXT:    ; implicit-def: $sgpr80
; GFX11-NEXT:    ; implicit-def: $sgpr70
; GFX11-NEXT:    ; implicit-def: $sgpr69
; GFX11-NEXT:    ; implicit-def: $sgpr68
; GFX11-NEXT:    ; implicit-def: $sgpr66
; GFX11-NEXT:    ; implicit-def: $sgpr58
; GFX11-NEXT:    ; implicit-def: $sgpr67
; GFX11-NEXT:    ; implicit-def: $sgpr65
; GFX11-NEXT:    ; implicit-def: $sgpr64
; GFX11-NEXT:    ; implicit-def: $sgpr55
; GFX11-NEXT:    ; implicit-def: $sgpr53
; GFX11-NEXT:    ; implicit-def: $sgpr60
; GFX11-NEXT:    ; implicit-def: $sgpr54
; GFX11-NEXT:    ; implicit-def: $sgpr52
; GFX11-NEXT:    ; implicit-def: $sgpr51
; GFX11-NEXT:    ; implicit-def: $sgpr50
; GFX11-NEXT:    ; implicit-def: $sgpr48
; GFX11-NEXT:    ; implicit-def: $sgpr49
; GFX11-NEXT:    ; implicit-def: $sgpr39
; GFX11-NEXT:    ; implicit-def: $sgpr38
; GFX11-NEXT:    ; implicit-def: $sgpr37
; GFX11-NEXT:    ; implicit-def: $sgpr35
; GFX11-NEXT:    ; implicit-def: $sgpr36
; GFX11-NEXT:    ; implicit-def: $sgpr34
; GFX11-NEXT:    ; implicit-def: $vcc_hi
; GFX11-NEXT:    ; implicit-def: $sgpr30
; GFX11-NEXT:    ; implicit-def: $sgpr94
; GFX11-NEXT:    ; implicit-def: $sgpr92
; GFX11-NEXT:    ; implicit-def: $sgpr90
; GFX11-NEXT:    ; implicit-def: $sgpr88
; GFX11-NEXT:    ; implicit-def: $sgpr78
; GFX11-NEXT:    ; implicit-def: $sgpr76
; GFX11-NEXT:    ; implicit-def: $sgpr74
; GFX11-NEXT:    ; implicit-def: $sgpr72
; GFX11-NEXT:    ; implicit-def: $sgpr62
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    s_branch .LBB37_2
; GFX11-NEXT:  .LBB37_4:
; GFX11-NEXT:    v_dual_mov_b32 v33, s0 :: v_dual_mov_b32 v34, s1
; GFX11-NEXT:    v_readlane_b32 s0, v76, 0
; GFX11-NEXT:    v_dual_mov_b32 v31, s2 :: v_dual_mov_b32 v32, s3
; GFX11-NEXT:    v_dual_mov_b32 v27, s16 :: v_dual_mov_b32 v28, s17
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v133, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 1
; GFX11-NEXT:    v_dual_mov_b32 v25, s18 :: v_dual_mov_b32 v26, s19
; GFX11-NEXT:    v_dual_mov_b32 v23, s20 :: v_dual_mov_b32 v24, s21
; GFX11-NEXT:    v_mov_b32_e32 v145, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 2
; GFX11-NEXT:    v_dual_mov_b32 v21, s22 :: v_dual_mov_b32 v22, s23
; GFX11-NEXT:    v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v131, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 3
; GFX11-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
; GFX11-NEXT:    v_mov_b32_e32 v135, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 4
; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
; GFX11-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v144, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 5
; GFX11-NEXT:    v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
; GFX11-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
; GFX11-NEXT:    v_mov_b32_e32 v132, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 6
; GFX11-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
; GFX11-NEXT:    v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v134, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 7
; GFX11-NEXT:    v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
; GFX11-NEXT:    v_dual_mov_b32 v72, s104 :: v_dual_mov_b32 v73, s102
; GFX11-NEXT:    v_mov_b32_e32 v128, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 8
; GFX11-NEXT:    v_dual_mov_b32 v63, s103 :: v_dual_mov_b32 v62, s100
; GFX11-NEXT:    v_dual_mov_b32 v61, s101 :: v_dual_mov_b32 v60, s97
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v129, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 9
; GFX11-NEXT:    v_dual_mov_b32 v57, s99 :: v_dual_mov_b32 v56, s98
; GFX11-NEXT:    v_dual_mov_b32 v58, s96 :: v_dual_mov_b32 v59, s87
; GFX11-NEXT:    v_mov_b32_e32 v130, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 10
; GFX11-NEXT:    v_dual_mov_b32 v46, s86 :: v_dual_mov_b32 v47, s84
; GFX11-NEXT:    v_dual_mov_b32 v42, s85 :: v_dual_mov_b32 v45, s82
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v115, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 11
; GFX11-NEXT:    v_dual_mov_b32 v44, s83 :: v_dual_mov_b32 v183, s81
; GFX11-NEXT:    v_dual_mov_b32 v43, s71 :: v_dual_mov_b32 v182, s80
; GFX11-NEXT:    v_mov_b32_e32 v119, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 12
; GFX11-NEXT:    v_dual_mov_b32 v40, s70 :: v_dual_mov_b32 v41, s69
; GFX11-NEXT:    v_dual_mov_b32 v178, s68 :: v_dual_mov_b32 v181, s66
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v113, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 13
; GFX11-NEXT:    v_dual_mov_b32 v176, s67 :: v_dual_mov_b32 v179, s65
; GFX11-NEXT:    v_dual_mov_b32 v180, s64 :: v_dual_mov_b32 v167, s55
; GFX11-NEXT:    v_mov_b32_e32 v117, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 14
; GFX11-NEXT:    v_dual_mov_b32 v177, s53 :: v_dual_mov_b32 v164, s54
; GFX11-NEXT:    v_dual_mov_b32 v165, s52 :: v_dual_mov_b32 v166, s51
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v118, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 15
; GFX11-NEXT:    v_dual_mov_b32 v160, s50 :: v_dual_mov_b32 v163, s48
; GFX11-NEXT:    v_dual_mov_b32 v150, s49 :: v_dual_mov_b32 v161, s39
; GFX11-NEXT:    v_mov_b32_e32 v114, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 16
; GFX11-NEXT:    v_dual_mov_b32 v162, s38 :: v_dual_mov_b32 v149, s37
; GFX11-NEXT:    v_dual_mov_b32 v151, s35 :: v_dual_mov_b32 v146, s36
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v116, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 17
; GFX11-NEXT:    v_dual_mov_b32 v147, s34 :: v_dual_mov_b32 v148, vcc_hi
; GFX11-NEXT:    v_dual_mov_b32 v29, s62 :: v_dual_mov_b32 v48, s88
; GFX11-NEXT:    v_mov_b32_e32 v102, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 18
; GFX11-NEXT:    v_dual_mov_b32 v35, s72 :: v_dual_mov_b32 v50, s92
; GFX11-NEXT:    v_dual_mov_b32 v37, s76 :: v_dual_mov_b32 v52, s30
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v103, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 19
; GFX11-NEXT:    v_dual_mov_b32 v49, s90 :: v_dual_mov_b32 v54, s58
; GFX11-NEXT:    v_dual_mov_b32 v51, s94 :: v_dual_mov_b32 v64, s56
; GFX11-NEXT:    v_mov_b32_e32 v112, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 20
; GFX11-NEXT:    v_dual_mov_b32 v53, s60 :: v_dual_mov_b32 v66, s44
; GFX11-NEXT:    v_mov_b32_e32 v65, s46
; GFX11-NEXT:    v_mov_b32_e32 v67, s42
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_mov_b32_e32 v97, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 21
; GFX11-NEXT:    v_mov_b32_e32 v36, s74
; GFX11-NEXT:    v_dual_mov_b32 v38, s78 :: v_dual_mov_b32 v101, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 22
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v87, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 23
; GFX11-NEXT:    v_mov_b32_e32 v99, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 24
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v100, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 25
; GFX11-NEXT:    v_mov_b32_e32 v96, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 26
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v98, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 27
; GFX11-NEXT:    v_mov_b32_e32 v80, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 28
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v85, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 29
; GFX11-NEXT:    v_mov_b32_e32 v86, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 30
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v82, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 31
; GFX11-NEXT:    v_mov_b32_e32 v84, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v55, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 1
; GFX11-NEXT:    v_mov_b32_e32 v81, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v83, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 3
; GFX11-NEXT:    v_mov_b32_e32 v30, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v71, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 5
; GFX11-NEXT:    v_mov_b32_e32 v39, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v69, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 7
; GFX11-NEXT:    v_mov_b32_e32 v70, s0
; GFX11-NEXT:  .LBB37_5: ; %end
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_perm_b32 v67, v73, v67, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v66, v60, v66, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v33, v33, v72, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v68, v61, v62, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v57, v31, v57, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 16, v67
; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
; GFX11-NEXT:    v_perm_b32 v58, v58, v59, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v34, v34, v63, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 16, v68
; GFX11-NEXT:    v_or_b32_e32 v31, v33, v67
; GFX11-NEXT:    v_perm_b32 v65, v47, v65, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v33, v57, v66
; GFX11-NEXT:    v_perm_b32 v66, v44, v45, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v64, v43, v64, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v67, v40, v41, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v56, v32, v56, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v58, 16, v58
; GFX11-NEXT:    v_perm_b32 v53, v177, v53, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v32, v34, v68
; GFX11-NEXT:    v_perm_b32 v27, v27, v46, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX11-NEXT:    v_perm_b32 v28, v28, v42, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v66, 16, v66
; GFX11-NEXT:    v_perm_b32 v68, v25, v183, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX11-NEXT:    v_perm_b32 v182, v26, v182, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v67, 16, v67
; GFX11-NEXT:    v_or_b32_e32 v34, v56, v58
; GFX11-NEXT:    v_perm_b32 v21, v21, v167, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX11-NEXT:    v_or_b32_e32 v25, v27, v65
; GFX11-NEXT:    v_or_b32_e32 v26, v28, v66
; GFX11-NEXT:    v_or_b32_e32 v27, v68, v64
; GFX11-NEXT:    v_or_b32_e32 v28, v182, v67
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_store_b128 v0, v[31:34], off
; GFX11-NEXT:    scratch_store_b128 v0, v[25:28], off offset:16
; GFX11-NEXT:    v_perm_b32 v26, v165, v166, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v25, v21, v53
; GFX11-NEXT:    v_perm_b32 v21, v22, v164, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v22, v163, v52, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v28, v151, v51, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v27, v161, v162, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-NEXT:    v_perm_b32 v19, v19, v160, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v17, v17, v149, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-NEXT:    v_perm_b32 v20, v20, v150, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-NEXT:    v_or_b32_e32 v26, v21, v26
; GFX11-NEXT:    v_or_b32_e32 v19, v19, v22
; GFX11-NEXT:    v_perm_b32 v22, v147, v148, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v21, v17, v28
; GFX11-NEXT:    v_perm_b32 v17, v18, v146, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v18, v145, v50, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v28, v134, v49, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v20, v20, v27
; GFX11-NEXT:    v_perm_b32 v27, v135, v144, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v15, v15, v133, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v13, v13, v132, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-NEXT:    v_perm_b32 v16, v16, v131, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-NEXT:    v_or_b32_e32 v22, v17, v22
; GFX11-NEXT:    v_or_b32_e32 v15, v15, v18
; GFX11-NEXT:    v_perm_b32 v18, v129, v130, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v17, v13, v28
; GFX11-NEXT:    v_perm_b32 v28, v116, v38, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v16, v16, v27
; GFX11-NEXT:    v_perm_b32 v13, v14, v128, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v14, v119, v48, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v27, v117, v118, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v9, v9, v114, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-NEXT:    v_perm_b32 v11, v11, v115, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v12, v12, v113, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-NEXT:    v_or_b32_e32 v18, v13, v18
; GFX11-NEXT:    v_or_b32_e32 v13, v9, v28
; GFX11-NEXT:    v_perm_b32 v9, v10, v102, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v10, v101, v37, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v11, v11, v14
; GFX11-NEXT:    v_perm_b32 v14, v103, v112, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v12, v12, v27
; GFX11-NEXT:    v_perm_b32 v27, v99, v100, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v28, v98, v36, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v7, v7, v97, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_perm_b32 v54, v181, v54, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v64, v179, v180, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v8, v8, v87, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-NEXT:    v_perm_b32 v5, v5, v96, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-NEXT:    v_or_b32_e32 v7, v7, v10
; GFX11-NEXT:    v_perm_b32 v10, v85, v86, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v23, v178, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX11-NEXT:    v_perm_b32 v24, v24, v176, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX11-NEXT:    v_or_b32_e32 v14, v9, v14
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v27
; GFX11-NEXT:    v_or_b32_e32 v9, v5, v28
; GFX11-NEXT:    v_perm_b32 v5, v84, v35, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v27, v81, v83, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v28, v71, v29, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v29, v69, v70, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v6, v6, v80, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_or_b32_e32 v23, v23, v54
; GFX11-NEXT:    v_or_b32_e32 v24, v24, v64
; GFX11-NEXT:    v_perm_b32 v3, v3, v82, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-NEXT:    v_perm_b32 v4, v4, v55, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-NEXT:    v_perm_b32 v30, v1, v30, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-NEXT:    v_perm_b32 v31, v2, v39, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-NEXT:    v_or_b32_e32 v10, v6, v10
; GFX11-NEXT:    v_or_b32_e32 v1, v3, v5
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v27
; GFX11-NEXT:    v_or_b32_e32 v3, v30, v28
; GFX11-NEXT:    v_or_b32_e32 v4, v31, v29
; GFX11-NEXT:    s_clause 0x5
; GFX11-NEXT:    scratch_store_b128 v0, v[23:26], off offset:32
; GFX11-NEXT:    scratch_store_b128 v0, v[19:22], off offset:48
; GFX11-NEXT:    scratch_store_b128 v0, v[15:18], off offset:64
; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v73, off, s32
; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-NEXT:    v_readlane_b32 s104, v75, 8
; GFX11-NEXT:    v_readlane_b32 s103, v75, 7
; GFX11-NEXT:    v_readlane_b32 s102, v75, 6
; GFX11-NEXT:    v_readlane_b32 s101, v75, 5
; GFX11-NEXT:    v_readlane_b32 s100, v75, 4
; GFX11-NEXT:    v_readlane_b32 s99, v75, 3
; GFX11-NEXT:    v_readlane_b32 s98, v75, 2
; GFX11-NEXT:    v_readlane_b32 s97, v75, 1
; GFX11-NEXT:    v_readlane_b32 s96, v75, 0
; GFX11-NEXT:    v_readlane_b32 s87, v74, 31
; GFX11-NEXT:    v_readlane_b32 s86, v74, 30
; GFX11-NEXT:    v_readlane_b32 s85, v74, 29
; GFX11-NEXT:    v_readlane_b32 s84, v74, 28
; GFX11-NEXT:    v_readlane_b32 s83, v74, 27
; GFX11-NEXT:    v_readlane_b32 s82, v74, 26
; GFX11-NEXT:    v_readlane_b32 s81, v74, 25
; GFX11-NEXT:    v_readlane_b32 s80, v74, 24
; GFX11-NEXT:    v_readlane_b32 s71, v74, 23
; GFX11-NEXT:    v_readlane_b32 s70, v74, 22
; GFX11-NEXT:    v_readlane_b32 s69, v74, 21
; GFX11-NEXT:    v_readlane_b32 s68, v74, 20
; GFX11-NEXT:    v_readlane_b32 s67, v74, 19
; GFX11-NEXT:    v_readlane_b32 s66, v74, 18
; GFX11-NEXT:    v_readlane_b32 s65, v74, 17
; GFX11-NEXT:    v_readlane_b32 s64, v74, 16
; GFX11-NEXT:    v_readlane_b32 s55, v74, 15
; GFX11-NEXT:    v_readlane_b32 s54, v74, 14
; GFX11-NEXT:    v_readlane_b32 s53, v74, 13
; GFX11-NEXT:    v_readlane_b32 s52, v74, 12
; GFX11-NEXT:    v_readlane_b32 s51, v74, 11
; GFX11-NEXT:    v_readlane_b32 s50, v74, 10
; GFX11-NEXT:    v_readlane_b32 s49, v74, 9
; GFX11-NEXT:    v_readlane_b32 s48, v74, 8
; GFX11-NEXT:    v_readlane_b32 s39, v74, 7
; GFX11-NEXT:    v_readlane_b32 s38, v74, 6
; GFX11-NEXT:    v_readlane_b32 s37, v74, 5
; GFX11-NEXT:    v_readlane_b32 s36, v74, 4
; GFX11-NEXT:    v_readlane_b32 s35, v74, 3
; GFX11-NEXT:    v_readlane_b32 s34, v74, 2
; GFX11-NEXT:    v_readlane_b32 s31, v74, 1
; GFX11-NEXT:    v_readlane_b32 s30, v74, 0
; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:72
; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:76
; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) {
; SI-LABEL: bitcast_v128i8_to_v32f32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:188
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:204
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:212
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:220
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:268
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:292
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:316
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:332
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:352
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:376
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:348
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:384
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380
; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:364
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:372
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB38_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    v_or_b32_e32 v2, v2, v41
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_or_b32_e32 v3, v40, v3
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v2, v2, v53
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v48
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v62
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v38
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v56
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v46
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
; SI-NEXT:    v_or_b32_e32 v8, v8, v44
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v52
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v43
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:  .LBB38_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB38_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_movk_i32 s6, 0x300
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b32 s7, 0x3000000
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v54, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v40, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v41, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v53, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v48, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v62, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v38, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v56, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v46, v7
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v8, v44, v8
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v33
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v59
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v18
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v27
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
; SI-NEXT:    v_or_b32_e32 v31, v43, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v31
; SI-NEXT:  .LBB38_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v32f32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:12
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:380
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB38_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s6, 0xc0c0004
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v9, v33, v62, s6
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_perm_b32 v10, v63, v60, s6
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v11, v53, v58, s6
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_perm_b32 v12, v51, v56, s6
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v37, v36, s6
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v45, v44, s6
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v49, v48, s6
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v0, v1, v0, s6
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s6
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v41, s6
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_perm_b32 v8, v55, v34, s6
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_perm_b32 v9, v61, v32, s6
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_perm_b32 v10, v59, v54, s6
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_perm_b32 v11, v57, v52, s6
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v47, v38, s6
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_perm_b32 v13, v35, v46, s6
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_perm_b32 v14, v43, v50, s6
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_perm_b32 v15, v39, v42, s6
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v30, v30, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v32, v40, v32, s6
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:  .LBB38_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB38_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v31, 0x300
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v41
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v62
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v60
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v58
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v56
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v36
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v48
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    s_waitcnt vmcnt(11)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    s_waitcnt vmcnt(10)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    s_waitcnt vmcnt(9)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    s_waitcnt vmcnt(8)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    s_waitcnt vmcnt(7)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    s_waitcnt vmcnt(6)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_add_u16_e32 v8, 3, v55
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v34
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 3, v33
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_add_u16_e32 v9, 3, v61
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v32
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 3, v63
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
; VI-NEXT:    v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_add_u16_e32 v10, 3, v59
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v54
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 3, v53
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_add_u16_e32 v11, 3, v57
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v52
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 3, v51
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_add_u16_e32 v12, 3, v47
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 3, v37
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
; VI-NEXT:    v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_add_u16_e32 v13, 3, v35
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v46
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 3, v45
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_add_u16_e32 v14, 3, v43
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v50
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v15, 3, v49
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_add_u16_e32 v15, 3, v39
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v42
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v30, v32
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v33, 3, v40
; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:  .LBB38_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v32f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:12
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:380
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB38_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0xc0c0004
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v10, v55, v32, s6
; GFX9-NEXT:    v_perm_b32 v11, v61, v60, s6
; GFX9-NEXT:    v_perm_b32 v12, v53, v38, s6
; GFX9-NEXT:    v_perm_b32 v13, v57, v56, s6
; GFX9-NEXT:    v_perm_b32 v14, v51, v34, s6
; GFX9-NEXT:    v_perm_b32 v15, v45, v50, s6
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v16, v43, v16, s6
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v9, s6
; GFX9-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX9-NEXT:    v_perm_b32 v8, v35, v42, s6
; GFX9-NEXT:    v_perm_b32 v9, v41, v48, s6
; GFX9-NEXT:    v_lshl_or_b32 v8, v9, 16, v8
; GFX9-NEXT:    v_perm_b32 v9, v39, v40, s6
; GFX9-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v10, v62, v63, s6
; GFX9-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX9-NEXT:    v_perm_b32 v11, v59, v54, s6
; GFX9-NEXT:    v_lshl_or_b32 v11, v12, 16, v11
; GFX9-NEXT:    v_perm_b32 v12, v58, v37, s6
; GFX9-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v47, v52, s6
; GFX9-NEXT:    v_lshl_or_b32 v13, v14, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v33, v46, s6
; GFX9-NEXT:    v_lshl_or_b32 v14, v15, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v49, v44, s6
; GFX9-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s6
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    v_lshl_or_b32 v16, v17, 16, v16
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    v_lshl_or_b32 v18, v19, 16, v18
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    v_lshl_or_b32 v19, v20, 16, v19
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    v_lshl_or_b32 v21, v22, 16, v21
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    v_lshl_or_b32 v23, v24, 16, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    v_lshl_or_b32 v24, v25, 16, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    v_lshl_or_b32 v25, v26, 16, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    v_lshl_or_b32 v26, v27, 16, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    v_lshl_or_b32 v28, v29, 16, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    v_lshl_or_b32 v29, v30, 16, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    v_lshl_or_b32 v30, v31, 16, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v32, v36, v32, s6
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    v_lshl_or_b32 v31, v32, 16, v31
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:  .LBB38_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB38_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    s_movk_i32 s6, 0x300
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v48
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v32
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v60
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v56
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v34
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v50
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(25)
; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(24)
; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
; GFX9-NEXT:    s_waitcnt vmcnt(22)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(19)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(14)
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(13)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    s_waitcnt vmcnt(11)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    s_waitcnt vmcnt(8)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v35
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v42
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v41
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v39
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v40
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v55
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v63
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v61
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v54
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v53
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v58
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v37
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v57
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v52
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v51
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v33
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v46
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v45
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v49
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v44
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v43
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v32, 3, v36
; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
; GFX9-NEXT:  .LBB38_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v32f32:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:384
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:380
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:376
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:372
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:368
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:364
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:360
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:356
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:352
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:348
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:344
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:340
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:336
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:332
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:328
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:260
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:132
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:388
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:8
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v74.l, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v75.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v76.l, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v77.l, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v78.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v88.l, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v79.l, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v89.l, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v90.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v92.l, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v91.l, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v93.l, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v94.l, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v95.l, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v104.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v106.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v105.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v107.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v108.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v110.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v109.l, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v111.l, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v120.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v121.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v122.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v123.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v124.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v125.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v126.l, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v127.l, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB38_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v127, v126, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v125, v124, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v123, v122, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v121, v120, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v107, v105, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v106, v104, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v95, v94, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v111, v109, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v110, v108, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v93, v91, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v92, v90, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v89, v79, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v88, v78, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v77, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v75, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v74, v72, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v63, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v62, v60, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v59, v58, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v57, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v56, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v45, v43, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v44, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v41, v40, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v183, v181, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v182, v180, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v179, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v178, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v164, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v161, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v160, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v149, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v147, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v146, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v135, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v134, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v131, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v129, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v128, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v117, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v116, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v113, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v103, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v102, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v99, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v98, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v87, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v84, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v81, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v80, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v69, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v66, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v55, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v54, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v48, v49, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v37, v39, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v38, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v36, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v48, 16, v31
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v35, 16, v37
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v32, 16, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr127_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr126_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr125_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr124_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr122_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr121_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr120_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr111_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr109_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr108_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr107_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr105_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr106_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr104_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr95_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr94_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr93_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr91_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr92_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr90_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr89_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr88_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT:  .LBB38_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB38_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v127.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v125.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v123.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v0.h, 8, v126.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v124.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v122.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v121.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v0.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v120.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v111.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v110.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v109.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v3.h, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v108.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, v107.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, v106.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v95.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v3.l, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v3.h, v2.h
; GFX11-TRUE16-NEXT:    v_and_b16 v3.l, 0xff, v4.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v105.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v4.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v104.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, v92.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v90.l
; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v3.h, v3.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, v89.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v94.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v93.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, v88.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v77.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v5.l, 8, v91.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, v74.l, 3
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v72.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, v63.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, v62.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v5.l, v4.h
; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v6.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v79.l
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v78.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v59.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, v56.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v76.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v75.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v10.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v46.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, v45.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.h, 8, v73.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, v44.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v41.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, v182.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v180.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v7.h, v7.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v9.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v61.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v9.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v60.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, v179.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, v178.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v58.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v57.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v167.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, v164.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v162.l
; GFX11-TRUE16-NEXT:    v_or_b16 v9.l, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v10.l, 8, v47.l
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, v161.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, v160.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v149.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v9.h, v10.l, v9.h
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v11.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v43.l
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v42.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, v146.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v144.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, v135.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.l, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v40.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v183.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, v134.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v131.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.h, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v12.l, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.h, 8, v181.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, v128.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v118.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, v117.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, v116.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.h, v12.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v14.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v177.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v14.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v176.l
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v113.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.l, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, v102.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.h, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v166.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v165.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v23.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v100.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, v99.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v15.l, 8, v163.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, v98.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v87.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, v84.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v82.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v15.l, v14.h
; GFX11-TRUE16-NEXT:    v_or_b16 v15.l, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v16.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v151.l
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v150.l
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, v81.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v15.h, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, v80.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v148.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v147.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v69.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, v66.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v28.h, 8, v64.l
; GFX11-TRUE16-NEXT:    v_or_b16 v16.h, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v17.l, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.h, 8, v145.l
; GFX11-TRUE16-NEXT:    v_and_b16 v28.l, 0xff, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, v55.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v29.h, 8, v53.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.l, 8, v51.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.l, v17.h, v17.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.h, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v19.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v133.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v19.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v132.l
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.h, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, v52.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v18.l, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v18.h, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v130.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v129.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v28.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v50.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.l, v29.h, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.l, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.h, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v20.l, 8, v119.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, v38.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, v36.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v33.h, v34.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.h, 8, v37.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.h, v20.l, v19.h
; GFX11-TRUE16-NEXT:    v_or_b16 v20.l, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v21.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v115.l
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v114.l
; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v31.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.l, 8, v35.l
; GFX11-TRUE16-NEXT:    v_or_b16 v20.h, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v21.l, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v112.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v103.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT:    v_and_b16 v33.h, 0xff, v33.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.l, 8, v32.l
; GFX11-TRUE16-NEXT:    v_or_b16 v21.h, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v22.l, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.h, 8, v101.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.l, v34.l, v31.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.h, v33.l, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v32.l, v33.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.l, v22.h, v22.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.h, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v24.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v97.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v24.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v96.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v23.l, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v23.h, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v86.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v85.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v25.l, 8, v83.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v3.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v4.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v4.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v5.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v25.l, v24.h
; GFX11-TRUE16-NEXT:    v_or_b16 v25.l, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v26.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v71.l
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v70.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v6.l
; GFX11-TRUE16-NEXT:    v_or_b16 v25.h, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v6.h
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v68.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v67.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v7.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v7.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v8.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.h, 8, v65.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v8.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v9.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v27.l, v27.h, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v54.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v11.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v11.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v12.l
; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v12.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v13.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v14.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.h, v30.l, v27.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v28.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, v49.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v29.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, v39.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v30.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v48.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v14.h
; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v15.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v15.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v16.l
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v32.h, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v16.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v17.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v17.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v18.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v19.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v19.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v20.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v21.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v21.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v22.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v22.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v23.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v24.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v24.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v25.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v26.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v26.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v30.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v31.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v31.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v32.l
; GFX11-TRUE16-NEXT:  .LBB38_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v32f32:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v86, v0
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:384
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:380
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:376
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:372
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:368
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:364
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:360
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:356
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:352
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:348
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:344
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:340
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:336
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:332
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:328
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:260
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:188
; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v45, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v46, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v47, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v57, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v56, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v59, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v61, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v62, off, s32 offset:132
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v63, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v73, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v72, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v74, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v75, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v79, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v89, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v88, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v90, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v91, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v93, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v94, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v95, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v105, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_u16 v104, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v106, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v107, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v109, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v108, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v110, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v111, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v120, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_u16 v121, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v122, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v0, off, s32 offset:388
; GFX11-FAKE16-NEXT:    scratch_load_u16 v123, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v125, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v124, off, s32 offset:8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_u16 v126, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v127, off, s32
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, v30 :: v_dual_mov_b32 v32, v29
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v28 :: v_dual_mov_b32 v35, v27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, v26 :: v_dual_mov_b32 v36, v25
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, v24 :: v_dual_mov_b32 v39, v23
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v49, v22 :: v_dual_mov_b32 v48, v21
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v16 :: v_dual_mov_b32 v55, v15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v65, v14 :: v_dual_mov_b32 v64, v13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, v12 :: v_dual_mov_b32 v67, v11
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, v10 :: v_dual_mov_b32 v68, v9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, v8 :: v_dual_mov_b32 v71, v7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, v4 :: v_dual_mov_b32 v83, v3
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, v2 :: v_dual_mov_b32 v84, v1
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB38_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v86, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v85, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v80, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v66, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v65, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v54, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v69, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v52, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v50, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v49, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v38, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v37, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v33, v127, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v126, v124, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v125, v123, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v122, v121, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v120, v111, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v110, v108, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v109, v107, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v106, v104, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v105, v95, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v94, v92, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v93, v91, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v90, v88, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v89, v79, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v78, v77, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v76, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v74, v72, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v73, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v62, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v61, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v58, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v57, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v46, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v45, v43, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v42, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v40, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v182, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v181, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v178, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v177, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v166, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v165, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v162, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v161, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v150, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v148, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v146, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v145, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v134, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v133, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v130, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v129, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v118, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v117, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v115, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v112, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v103, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v34, v101, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v99, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v97, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v31
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v34, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v36, 16, v35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr127
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr126
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr124
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr125
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr122
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr121
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr120
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr111
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr108
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr109
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr107
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr106
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr105
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr95
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr93
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr91
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr89
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:  .LBB38_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB38_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v86, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v85, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v84
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v82, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v80, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v81
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v69, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v70, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v67
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v68
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v66, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v54, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v65, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v55
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v7, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v52, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v51
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v50, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v49, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v8, 8, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, v38, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, v37, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v34, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v33, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v10, 8, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v127
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v126, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v125, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v11, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v124
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v123
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v122, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v121
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v120, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v111
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v110, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v109, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v107
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v13, 8, v108
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, v106, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, v105, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v14, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v104
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v95
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v94, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v12, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v93, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v15, 8, v92
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v91
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v90, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v89, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v88
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v79
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v78, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v17, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v77
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v76, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v14, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v75
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v74, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v73, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v63
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v18, 8, v72
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, v62, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, v61, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v19, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v60
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v59
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v58, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v57, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v20, 8, v56
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v47
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v46, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v45, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v21, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v44
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v43
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v42, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v22, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v41
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v40, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v183
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v182, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v181, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v179
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v23, 8, v180
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, v178, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, v177, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v24, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v176
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v167
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v166, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v22, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v165, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v25, 8, v164
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v163
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v162, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v161, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v26, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v160
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v151
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v150, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v27, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v149
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v148, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v146, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v145, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v135
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v28, 8, v144
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v134, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v133, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v29, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v132
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v131
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, v130, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v129, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v30, 8, v128
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v119
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v26, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v118, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v117, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v31, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v115, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v116
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v33, 8, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v113
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v32, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, v112, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, v103, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, v101, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, v99, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v38, v97, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v34, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v102
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v37, 8, v100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v39, 8, v98
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v48, 8, v96
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v49, 8, v87
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v34, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v37, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v39, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v48, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v49, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v35, v36
; GFX11-FAKE16-NEXT:  .LBB38_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define inreg <32 x float> @bitcast_v128i8_to_v32f32_scalar(<128 x i8> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v128i8_to_v32f32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s76, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s46, v30
; SI-NEXT:    v_readfirstlane_b32 s6, v29
; SI-NEXT:    v_readfirstlane_b32 s7, v28
; SI-NEXT:    v_readfirstlane_b32 s8, v27
; SI-NEXT:    v_readfirstlane_b32 s9, v26
; SI-NEXT:    v_readfirstlane_b32 s10, v25
; SI-NEXT:    v_readfirstlane_b32 s11, v24
; SI-NEXT:    v_readfirstlane_b32 s12, v23
; SI-NEXT:    v_readfirstlane_b32 s13, v22
; SI-NEXT:    v_readfirstlane_b32 s14, v21
; SI-NEXT:    v_readfirstlane_b32 s15, v20
; SI-NEXT:    v_readfirstlane_b32 s40, v19
; SI-NEXT:    v_readfirstlane_b32 s41, v18
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:96
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:104
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:328
; SI-NEXT:    v_readfirstlane_b32 s75, v1
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:332
; SI-NEXT:    v_readfirstlane_b32 s43, v16
; SI-NEXT:    v_readfirstlane_b32 s45, v14
; SI-NEXT:    v_readfirstlane_b32 s73, v3
; SI-NEXT:    v_readfirstlane_b32 s74, v2
; SI-NEXT:    v_readfirstlane_b32 s57, v11
; SI-NEXT:    v_readfirstlane_b32 s72, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v15
; SI-NEXT:    v_readfirstlane_b32 s47, v13
; SI-NEXT:    v_readfirstlane_b32 s56, v12
; SI-NEXT:    v_readfirstlane_b32 s58, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s60, v8
; SI-NEXT:    v_readfirstlane_b32 s61, v7
; SI-NEXT:    v_readfirstlane_b32 s62, v6
; SI-NEXT:    v_readfirstlane_b32 s63, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v50
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v51
; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v52
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v13, 24, v54
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v55
; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v40
; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v41
; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v42
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v43
; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v46
; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v56
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v59
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v61
; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v44
; SI-NEXT:    v_lshlrev_b32_e32 v52, 8, v45
; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v47
; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:316
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v30
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 24, v29
; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v28
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v26
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v25
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v24
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v23
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v22
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v20
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v19
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v18
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:268
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:260
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:252
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:292
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v34
; SI-NEXT:    v_lshlrev_b32_e32 v34, 8, v32
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:220
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204
; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 24, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v33
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:196
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:116
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB39_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v31
; SI-NEXT:    v_or_b32_e32 v0, s4, v0
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v12
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v11, v3
; SI-NEXT:    v_or_b32_e32 v11, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v30
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v13, v1
; SI-NEXT:    v_or_b32_e32 v13, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v9
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v14, v1
; SI-NEXT:    v_or_b32_e32 v14, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v42
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v0, v0, v60
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v15, v1
; SI-NEXT:    v_or_b32_e32 v15, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v63
; SI-NEXT:    v_or_b32_e32 v0, v0, v41
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v16, v1
; SI-NEXT:    v_or_b32_e32 v16, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v61
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v17
; SI-NEXT:    v_or_b32_e32 v0, v0, v40
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v55, v1
; SI-NEXT:    v_or_b32_e32 v17, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v44
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v53, v1
; SI-NEXT:    v_or_b32_e32 v18, v0, v1
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v51, v1
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v12, v2, v3
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s18, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s77, s19, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s77, s5
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s77, s21, 8
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s22, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s77, s78, s77
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s24, 0xff
; SI-NEXT:    s_lshl_b32 s78, s25, 8
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s26, 0xff
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s79, s27, 24
; SI-NEXT:    s_and_b32 s77, s77, 0xffff
; SI-NEXT:    s_or_b32 s78, s79, s78
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s28, 0xff
; SI-NEXT:    s_lshl_b32 s79, s29, 8
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s76, 0xff
; SI-NEXT:    s_lshl_b32 s79, s79, 16
; SI-NEXT:    s_lshl_b32 s88, s75, 24
; SI-NEXT:    s_and_b32 s78, s78, 0xffff
; SI-NEXT:    s_or_b32 s79, s88, s79
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s74, 0xff
; SI-NEXT:    s_lshl_b32 s88, s73, 8
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s72, 0xff
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s89, s63, 24
; SI-NEXT:    s_and_b32 s79, s79, 0xffff
; SI-NEXT:    s_or_b32 s88, s89, s88
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s62, 0xff
; SI-NEXT:    s_lshl_b32 s89, s61, 8
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s60, 0xff
; SI-NEXT:    s_lshl_b32 s89, s89, 16
; SI-NEXT:    s_lshl_b32 s90, s59, 24
; SI-NEXT:    s_and_b32 s88, s88, 0xffff
; SI-NEXT:    s_or_b32 s89, s90, s89
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s58, 0xff
; SI-NEXT:    s_lshl_b32 s90, s57, 8
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s56, 0xff
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_lshl_b32 s91, s47, 24
; SI-NEXT:    s_and_b32 s89, s89, 0xffff
; SI-NEXT:    s_or_b32 s90, s91, s90
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s45, 0xff
; SI-NEXT:    s_lshl_b32 s91, s44, 8
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s43, 0xff
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v52
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v19, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v49, v1
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    s_lshl_b32 s91, s91, 16
; SI-NEXT:    s_lshl_b32 s92, s42, 24
; SI-NEXT:    s_and_b32 s90, s90, 0xffff
; SI-NEXT:    s_or_b32 s91, s92, s91
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s41, 0xff
; SI-NEXT:    s_lshl_b32 s92, s40, 8
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s15, 0xff
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s93, s14, 24
; SI-NEXT:    s_and_b32 s91, s91, 0xffff
; SI-NEXT:    s_or_b32 s92, s93, s92
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s13, 0xff
; SI-NEXT:    s_lshl_b32 s93, s12, 8
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s11, 0xff
; SI-NEXT:    s_lshl_b32 s93, s93, 16
; SI-NEXT:    s_lshl_b32 s94, s10, 24
; SI-NEXT:    s_and_b32 s92, s92, 0xffff
; SI-NEXT:    s_or_b32 s93, s94, s93
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s9, 0xff
; SI-NEXT:    s_lshl_b32 s94, s8, 8
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    s_and_b32 s94, s7, 0xff
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_lshl_b32 s95, s6, 24
; SI-NEXT:    s_and_b32 s93, s93, 0xffff
; SI-NEXT:    s_or_b32 s94, s95, s94
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v3, s78
; SI-NEXT:    v_mov_b32_e32 v4, s79
; SI-NEXT:    v_mov_b32_e32 v5, s88
; SI-NEXT:    v_mov_b32_e32 v6, s89
; SI-NEXT:    v_mov_b32_e32 v7, s90
; SI-NEXT:    v_mov_b32_e32 v8, s91
; SI-NEXT:    v_mov_b32_e32 v9, s92
; SI-NEXT:    v_mov_b32_e32 v10, s93
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v52
; SI-NEXT:    v_or_b32_e32 v0, v0, v50
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v20, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v39, v1
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v48
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v21, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v37, v1
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v48
; SI-NEXT:    v_or_b32_e32 v0, v0, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v22, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v35, v1
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v38
; SI-NEXT:    v_or_b32_e32 v0, v0, v36
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v23, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v33, v1
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v36
; SI-NEXT:    v_or_b32_e32 v0, v0, v34
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v24, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    v_or_b32_e32 v0, v0, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v29, v1
; SI-NEXT:    v_or_b32_e32 v25, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v28, v1
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v0, v0, v59
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v26, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v58
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v56
; SI-NEXT:    v_or_b32_e32 v0, v0, v57
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v27, v1
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    v_or_b32_e32 v27, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v0, v0, v45
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_or_b32_e32 v28, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v46
; SI-NEXT:    v_mov_b32_e32 v45, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v60, v32
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v29, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v30, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v31, v0, v1
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s77
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_branch .LBB39_3
; SI-NEXT:  .LBB39_2:
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB39_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v43, v58
; SI-NEXT:    v_mov_b32_e32 v58, v33
; SI-NEXT:    v_mov_b32_e32 v33, v37
; SI-NEXT:    v_mov_b32_e32 v34, v39
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_cbranch_vccnz .LBB39_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_mov_b32_e32 v32, v35
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_and_b32 s17, s22, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s76, s76, 3
; SI-NEXT:    s_add_i32 s74, s74, 3
; SI-NEXT:    s_add_i32 s72, s72, 3
; SI-NEXT:    s_add_i32 s62, s62, 3
; SI-NEXT:    s_add_i32 s60, s60, 3
; SI-NEXT:    s_add_i32 s58, s58, 3
; SI-NEXT:    s_add_i32 s56, s56, 3
; SI-NEXT:    s_and_b32 s22, s56, 0xff
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_addk_i32 s12, 0x300
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    s_add_i32 s10, s10, 0x3000000
; SI-NEXT:    s_add_i32 s6, s6, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v9, s10
; SI-NEXT:    v_mov_b32_e32 v10, s6
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, s4, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, 0x3000000, v0
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s16, s18, 0xff
; SI-NEXT:    s_addk_i32 s4, 0x300
; SI-NEXT:    s_lshl_b32 s5, s19, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s5, s16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s16, s21, 8
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_addk_i32 s5, 0x300
; SI-NEXT:    s_lshl_b32 s16, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_and_b32 s16, s24, 0xff
; SI-NEXT:    s_lshl_b32 s17, s25, 8
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s18, s26, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s27, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s28, 0xff
; SI-NEXT:    s_lshl_b32 s18, s29, 8
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s19, s76, 0xff
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_lshl_b32 s18, s75, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s18, s74, 0xff
; SI-NEXT:    s_lshl_b32 s19, s73, 8
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s20, s72, 0xff
; SI-NEXT:    s_addk_i32 s18, 0x300
; SI-NEXT:    s_lshl_b32 s19, s63, 24
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s19, s19, s20
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s19, s62, 0xff
; SI-NEXT:    s_lshl_b32 s20, s61, 8
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s21, s60, 0xff
; SI-NEXT:    s_addk_i32 s19, 0x300
; SI-NEXT:    s_lshl_b32 s20, s59, 24
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s21
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s20, s58, 0xff
; SI-NEXT:    s_lshl_b32 s21, s57, 8
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_addk_i32 s20, 0x300
; SI-NEXT:    s_lshl_b32 s21, s47, 24
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_and_b32 s21, s45, 0xff
; SI-NEXT:    s_lshl_b32 s22, s44, 8
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    s_and_b32 s23, s43, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s22, s42, 24
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s23
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s22, s41, 0xff
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_lshl_b32 s23, s40, 8
; SI-NEXT:    s_or_b32 s22, s23, s22
; SI-NEXT:    s_addk_i32 s22, 0x300
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s14, s14, s22
; SI-NEXT:    s_add_i32 s4, s4, 0x3000000
; SI-NEXT:    s_add_i32 s5, s5, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 0x3000000
; SI-NEXT:    s_add_i32 s17, s17, 0x3000000
; SI-NEXT:    s_add_i32 s18, s18, 0x3000000
; SI-NEXT:    s_add_i32 s19, s19, 0x3000000
; SI-NEXT:    s_add_i32 s20, s20, 0x3000000
; SI-NEXT:    s_add_i32 s21, s21, 0x3000000
; SI-NEXT:    s_add_i32 s14, s14, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v4, s18
; SI-NEXT:    v_mov_b32_e32 v5, s19
; SI-NEXT:    v_mov_b32_e32 v6, s20
; SI-NEXT:    v_mov_b32_e32 v7, s21
; SI-NEXT:    v_mov_b32_e32 v8, s14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_add_i32_e32 v12, vcc, 0x3000000, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v39, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_mov_b32_e32 v3, s17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v13, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v37, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v42, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v41, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v44
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v55, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v54
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v53, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v51, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v48
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v34, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v38
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v33, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v32, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v58
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v62, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v60
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v59, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v26, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v57, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v27, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v45, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v28, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v0
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:  .LBB39_5: ; %end
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v32f32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s45, v15
; VI-NEXT:    v_readfirstlane_b32 s46, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s76, v0
; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_readfirstlane_b32 s6, v30
; VI-NEXT:    v_readfirstlane_b32 s7, v29
; VI-NEXT:    v_readfirstlane_b32 s8, v28
; VI-NEXT:    v_readfirstlane_b32 s9, v27
; VI-NEXT:    v_readfirstlane_b32 s10, v26
; VI-NEXT:    v_readfirstlane_b32 s11, v25
; VI-NEXT:    v_readfirstlane_b32 s12, v24
; VI-NEXT:    v_readfirstlane_b32 s13, v23
; VI-NEXT:    v_readfirstlane_b32 s14, v22
; VI-NEXT:    v_readfirstlane_b32 s15, v21
; VI-NEXT:    v_readfirstlane_b32 s40, v20
; VI-NEXT:    v_readfirstlane_b32 s41, v19
; VI-NEXT:    v_readfirstlane_b32 s42, v18
; VI-NEXT:    v_readfirstlane_b32 s43, v17
; VI-NEXT:    v_readfirstlane_b32 s44, v16
; VI-NEXT:    v_readfirstlane_b32 s56, v12
; VI-NEXT:    v_readfirstlane_b32 s57, v11
; VI-NEXT:    v_readfirstlane_b32 s58, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s60, v8
; VI-NEXT:    v_readfirstlane_b32 s61, v7
; VI-NEXT:    v_readfirstlane_b32 s62, v6
; VI-NEXT:    v_readfirstlane_b32 s63, v5
; VI-NEXT:    v_readfirstlane_b32 s72, v4
; VI-NEXT:    v_readfirstlane_b32 s73, v3
; VI-NEXT:    v_readfirstlane_b32 s74, v2
; VI-NEXT:    v_readfirstlane_b32 s75, v1
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; VI-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; VI-NEXT:    s_and_b64 s[4:5], vcc, exec
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    s_cbranch_scc0 .LBB39_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_mov_b32_e32 v3, s23
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_mov_b32_e32 v2, s21
; VI-NEXT:    v_perm_b32 v3, s22, v3, v11
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_perm_b32 v2, s20, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_or_b32_e32 v10, v10, v12
; VI-NEXT:    v_perm_b32 v12, v14, v13, s4
; VI-NEXT:    v_perm_b32 v11, s6, v15, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v52, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v54, v53, s4
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    v_perm_b32 v13, v48, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v50, v49, s4
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_perm_b32 v14, v36, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v38, v37, s4
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_perm_b32 v15, v32, v16, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v16, v34, v33, s4
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    v_perm_b32 v16, v60, v17, s4
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_perm_b32 v17, v62, v61, s4
; VI-NEXT:    v_or_b32_e32 v16, v17, v16
; VI-NEXT:    v_perm_b32 v17, v59, v18, s4
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_perm_b32 v18, v41, v40, s4
; VI-NEXT:    v_or_b32_e32 v17, v18, v17
; VI-NEXT:    v_perm_b32 v18, v57, v19, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v19, v42, v58, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_perm_b32 v19, v45, v20, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_perm_b32 v20, v56, v47, s4
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_perm_b32 v20, v44, v21, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_perm_b32 v21, v46, v26, s4
; VI-NEXT:    v_or_b32_e32 v20, v21, v20
; VI-NEXT:    v_perm_b32 v21, v31, v22, s4
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_perm_b32 v22, v24, v43, s4
; VI-NEXT:    v_or_b32_e32 v21, v22, v21
; VI-NEXT:    v_perm_b32 v22, v28, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_perm_b32 v23, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v22, v23, v22
; VI-NEXT:    v_perm_b32 v23, v25, v63, s4
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_perm_b32 v24, v27, v55, s4
; VI-NEXT:    v_or_b32_e32 v23, v24, v23
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    v_or_b32_e32 v24, v25, v24
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    v_or_b32_e32 v25, v26, v25
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    v_or_b32_e32 v26, v27, v26
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    v_or_b32_e32 v27, v28, v27
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v28, v29, v28
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    v_or_b32_e32 v29, v30, v29
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    v_or_b32_e32 v30, v31, v30
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v43, v46, v43, s4
; VI-NEXT:    v_or_b32_e32 v31, v43, v31
; VI-NEXT:    s_mov_b64 s[4:5], 0
; VI-NEXT:    s_branch .LBB39_3
; VI-NEXT:  .LBB39_2:
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    s_mov_b64 s[4:5], -1
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:  .LBB39_3: ; %Flow
; VI-NEXT:    v_mov_b32_e32 v43, v44
; VI-NEXT:    v_mov_b32_e32 v44, v45
; VI-NEXT:    v_mov_b32_e32 v45, v47
; VI-NEXT:    v_mov_b32_e32 v46, v57
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; VI-NEXT:    s_cbranch_vccnz .LBB39_5
; VI-NEXT:  ; %bb.4: ; %cmp.true
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    v_mov_b32_e32 v1, s21
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    v_mov_b32_e32 v2, s23
; VI-NEXT:    v_perm_b32 v1, s20, v1, v11
; VI-NEXT:    v_perm_b32 v2, s22, v2, v11
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x300, v2
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    s_add_i32 s76, s76, 3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x300, v3
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s74, s74, 3
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    s_add_i32 s72, s72, 3
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x300, v4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s62, s62, 3
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    s_add_i32 s60, s60, 3
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x300, v5
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s58, s58, 3
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    s_add_i32 s56, s56, 3
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x300, v6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x300, v7
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x300, v8
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x300, v9
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x300, v10
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v10, v12, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v48
; VI-NEXT:    v_perm_b32 v14, v14, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v36
; VI-NEXT:    v_perm_b32 v15, v15, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x3000000, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x3000000, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x3000000, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v6
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v7
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x3000000, v8
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x3000000, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v10
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v11, s6, v12, v11
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x300, v11
; VI-NEXT:    v_perm_b32 v16, v16, v17, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v60
; VI-NEXT:    v_perm_b32 v17, v17, v18, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v57
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v18, v18, v19, s4
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v12, v12, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v54
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v52
; VI-NEXT:    v_perm_b32 v12, v12, v53, s4
; VI-NEXT:    v_perm_b32 v13, v13, v51, s4
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x300, v12
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v50
; VI-NEXT:    v_perm_b32 v13, v13, v49, s4
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x300, v13
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v38
; VI-NEXT:    v_perm_b32 v14, v14, v37, s4
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x300, v14
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v34
; VI-NEXT:    v_perm_b32 v15, v15, v33, s4
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x300, v15
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v62
; VI-NEXT:    v_perm_b32 v16, v16, v61, s4
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x300, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v41
; VI-NEXT:    v_perm_b32 v17, v17, v40, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x300, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v42
; VI-NEXT:    v_perm_b32 v18, v18, v47, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v18
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v46
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x3000000, v11
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x3000000, v12
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x3000000, v13
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x3000000, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x3000000, v15
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x3000000, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x3000000, v17
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x3000000, v18
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_perm_b32 v20, v20, v21, s4
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v45
; VI-NEXT:    v_perm_b32 v20, v20, v44, s4
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x300, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x3000000, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v43
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x3000000, v21
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x3000000, v22
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x3000000, v23
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x3000000, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x3000000, v25
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x3000000, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x3000000, v27
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x3000000, v28
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x3000000, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x3000000, v30
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x300, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v32
; VI-NEXT:    v_perm_b32 v32, v32, v33, s4
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x3000000, v31
; VI-NEXT:  .LBB39_5: ; %end
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v32f32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s45, v15
; GFX9-NEXT:    v_readfirstlane_b32 s46, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s76, v0
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readfirstlane_b32 s6, v30
; GFX9-NEXT:    v_readfirstlane_b32 s7, v29
; GFX9-NEXT:    v_readfirstlane_b32 s8, v28
; GFX9-NEXT:    v_readfirstlane_b32 s9, v27
; GFX9-NEXT:    v_readfirstlane_b32 s10, v26
; GFX9-NEXT:    v_readfirstlane_b32 s11, v25
; GFX9-NEXT:    v_readfirstlane_b32 s12, v24
; GFX9-NEXT:    v_readfirstlane_b32 s13, v23
; GFX9-NEXT:    v_readfirstlane_b32 s14, v22
; GFX9-NEXT:    v_readfirstlane_b32 s15, v21
; GFX9-NEXT:    v_readfirstlane_b32 s40, v20
; GFX9-NEXT:    v_readfirstlane_b32 s41, v19
; GFX9-NEXT:    v_readfirstlane_b32 s42, v18
; GFX9-NEXT:    v_readfirstlane_b32 s43, v17
; GFX9-NEXT:    v_readfirstlane_b32 s44, v16
; GFX9-NEXT:    v_readfirstlane_b32 s56, v12
; GFX9-NEXT:    v_readfirstlane_b32 s57, v11
; GFX9-NEXT:    v_readfirstlane_b32 s58, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s60, v8
; GFX9-NEXT:    v_readfirstlane_b32 s61, v7
; GFX9-NEXT:    v_readfirstlane_b32 s62, v6
; GFX9-NEXT:    v_readfirstlane_b32 s63, v5
; GFX9-NEXT:    v_readfirstlane_b32 s72, v4
; GFX9-NEXT:    v_readfirstlane_b32 s73, v3
; GFX9-NEXT:    v_readfirstlane_b32 s74, v2
; GFX9-NEXT:    v_readfirstlane_b32 s75, v1
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    s_cbranch_scc0 .LBB39_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v12
; GFX9-NEXT:    v_perm_b32 v12, v14, v13, s4
; GFX9-NEXT:    v_perm_b32 v11, s6, v15, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_perm_b32 v12, v52, v51, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v54, v53, s4
; GFX9-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX9-NEXT:    v_perm_b32 v13, v48, v39, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v50, v49, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX9-NEXT:    v_perm_b32 v14, v36, v35, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v38, v37, s4
; GFX9-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX9-NEXT:    v_perm_b32 v15, v32, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    v_perm_b32 v16, v34, v33, s4
; GFX9-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX9-NEXT:    v_perm_b32 v16, v60, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_perm_b32 v17, v62, v61, s4
; GFX9-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX9-NEXT:    v_perm_b32 v17, v59, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_perm_b32 v18, v41, v40, s4
; GFX9-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX9-NEXT:    v_perm_b32 v18, v57, v19, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_perm_b32 v19, v42, v58, s4
; GFX9-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX9-NEXT:    v_perm_b32 v19, v45, v20, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX9-NEXT:    v_perm_b32 v20, v56, v47, s4
; GFX9-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX9-NEXT:    v_perm_b32 v20, v44, v21, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX9-NEXT:    v_perm_b32 v21, v46, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX9-NEXT:    v_perm_b32 v21, v31, v22, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX9-NEXT:    v_perm_b32 v22, v24, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX9-NEXT:    v_perm_b32 v22, v28, v23, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX9-NEXT:    v_perm_b32 v23, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX9-NEXT:    v_perm_b32 v23, v25, v63, s4
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX9-NEXT:    v_perm_b32 v24, v27, v55, s4
; GFX9-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s4
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    v_or_b32_e32 v30, v31, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v43, v46, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v31, v43, v31
; GFX9-NEXT:    s_mov_b64 s[4:5], 0
; GFX9-NEXT:    s_branch .LBB39_3
; GFX9-NEXT:  .LBB39_2:
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    s_mov_b64 s[4:5], -1
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:  .LBB39_3: ; %Flow
; GFX9-NEXT:    v_mov_b32_e32 v43, v44
; GFX9-NEXT:    v_mov_b32_e32 v44, v45
; GFX9-NEXT:    v_mov_b32_e32 v45, v47
; GFX9-NEXT:    v_mov_b32_e32 v46, v57
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_vccnz .LBB39_5
; GFX9-NEXT:  ; %bb.4: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    s_movk_i32 s4, 0x300
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    s_add_i32 s76, s76, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_add_u32_e32 v3, 0x300, v3
; GFX9-NEXT:    v_add_u32_sdwa v4, v4, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s74, s74, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    s_add_i32 s72, s72, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_add_u32_e32 v4, 0x300, v4
; GFX9-NEXT:    v_add_u32_sdwa v5, v5, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s62, s62, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    s_add_i32 s60, s60, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_add_u32_e32 v5, 0x300, v5
; GFX9-NEXT:    v_add_u32_sdwa v6, v6, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s58, s58, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    s_add_i32 s56, s56, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_add_u32_e32 v6, 0x300, v6
; GFX9-NEXT:    v_add_u32_sdwa v7, v7, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_add_u32_e32 v7, 0x300, v7
; GFX9-NEXT:    v_add_u32_sdwa v8, v8, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_add_u32_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u32_sdwa v9, v9, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_add_u32_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u32_sdwa v10, v10, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s5, 0xc0c0004
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v32
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v48
; GFX9-NEXT:    v_perm_b32 v14, v14, v39, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v14, v14, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v36
; GFX9-NEXT:    v_perm_b32 v15, v15, v35, s5
; GFX9-NEXT:    v_add_u32_sdwa v15, v15, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v60
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_perm_b32 v17, v17, v18, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v57
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v18, v18, v19, s5
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v11, s6, v12, v11
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u32_sdwa v16, v16, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v17, v17, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v18, v18, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_perm_b32 v12, v12, v13, s5
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v54
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v52
; GFX9-NEXT:    v_perm_b32 v12, v12, v53, s5
; GFX9-NEXT:    v_perm_b32 v13, v13, v51, s5
; GFX9-NEXT:    v_add_u32_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u32_sdwa v13, v13, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v50
; GFX9-NEXT:    v_perm_b32 v13, v13, v49, s5
; GFX9-NEXT:    v_add_u32_e32 v13, 0x300, v13
; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v38
; GFX9-NEXT:    v_perm_b32 v14, v14, v37, s5
; GFX9-NEXT:    v_add_u32_e32 v14, 0x300, v14
; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v34
; GFX9-NEXT:    v_perm_b32 v15, v15, v33, s5
; GFX9-NEXT:    v_add_u32_e32 v15, 0x300, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v62
; GFX9-NEXT:    v_perm_b32 v16, v16, v61, s5
; GFX9-NEXT:    v_add_u32_e32 v16, 0x300, v16
; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v41
; GFX9-NEXT:    v_perm_b32 v17, v17, v40, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 0x300, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v42
; GFX9-NEXT:    v_perm_b32 v18, v18, v47, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 0x300, v18
; GFX9-NEXT:    v_add_u32_sdwa v19, v19, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v46
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_perm_b32 v20, v20, v21, s5
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v20, v20, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v45
; GFX9-NEXT:    v_perm_b32 v20, v20, v44, s5
; GFX9-NEXT:    v_add_u32_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v21, v21, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    v_add_u32_sdwa v22, v22, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    v_add_u32_sdwa v23, v23, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    v_add_u32_sdwa v24, v24, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    v_add_u32_sdwa v25, v25, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    v_add_u32_sdwa v26, v26, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    v_add_u32_sdwa v27, v27, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    v_add_u32_sdwa v28, v28, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    v_add_u32_sdwa v29, v29, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    v_add_u32_sdwa v30, v30, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    v_add_u32_sdwa v31, v31, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v32, 3, v32
; GFX9-NEXT:    v_perm_b32 v32, v32, v33, s5
; GFX9-NEXT:    v_add_u32_sdwa v32, v32, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:  .LBB39_5: ; %end
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v32f32_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:260
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:188
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:60
; GFX11-TRUE16-NEXT:    s_clause 0xf
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s75, 0
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB39_4
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v135, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v55, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v147, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v99, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v103, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v67, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB39_3
; GFX11-TRUE16-NEXT:  .LBB39_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v135
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v99
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v55
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-TRUE16-NEXT:  .LBB39_3: ; %end
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT:  .LBB39_4:
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_branch .LBB39_2
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v32f32_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v32, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v34, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v33, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v37, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v35, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v39, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v36, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v48, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v38, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v49, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v50, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v52, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v51, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:260
; GFX11-FAKE16-NEXT:    scratch_load_u16 v53, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v54, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v55, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:188
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:132
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:60
; GFX11-FAKE16-NEXT:    s_clause 0xf
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s75, 0
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-FAKE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB39_4
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v144, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v64, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v147, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v103, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v67, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB39_3
; GFX11-FAKE16-NEXT:  .LBB39_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v144
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v100
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v64
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-FAKE16-NEXT:  .LBB39_3: ; %end
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-FAKE16-NEXT:  .LBB39_4:
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_branch .LBB39_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define <64 x bfloat> @bitcast_v32f32_to_v64bf16(<32 x float> %a, i32 %b) {
; SI-LABEL: bitcast_v32f32_to_v64bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB40_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v62
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v62
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v13
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v11
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v10
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v9
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v8
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v7
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v3
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v2
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:  .LBB40_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB40_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_f32_e32 v31, 1.0, v62
; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v13
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v11
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v10
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v9
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v8
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v7
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v3
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v2
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:  .LBB40_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v60
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v58
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v56
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v46
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v44
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v42
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v40
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v54
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v52
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v50
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v48
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v39
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v38
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v37
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v36
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v35
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v34
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v33
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v32
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_alignbit_b32 v18, v18, v19, 16
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_alignbit_b32 v19, v19, v20, 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_alignbit_b32 v21, v21, v22, 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_alignbit_b32 v22, v22, v23, 16
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_alignbit_b32 v23, v23, v24, 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_alignbit_b32 v24, v24, v25, 16
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_alignbit_b32 v25, v25, v26, 16
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_alignbit_b32 v26, v26, v27, 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_alignbit_b32 v27, v27, v28, 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_alignbit_b32 v28, v28, v29, 16
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_alignbit_b32 v29, v29, v30, 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_alignbit_b32 v30, v30, v31, 16
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v63
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v64bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB40_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; VI-NEXT:  .LBB40_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v64bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB40_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15
; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14
; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13
; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12
; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11
; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10
; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9
; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8
; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7
; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6
; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5
; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4
; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3
; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2
; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30
; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29
; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28
; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27
; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26
; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25
; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24
; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23
; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22
; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21
; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20
; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19
; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18
; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17
; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16
; GFX9-NEXT:  .LBB40_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v64bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB40_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
; GFX11-NEXT:  .LBB40_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define inreg <64 x bfloat> @bitcast_v32f32_to_v64bf16_scalar(<32 x float> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32f32_to_v64bf16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_writelane_b32 v63, s70, 22
; SI-NEXT:    v_writelane_b32 v63, s71, 23
; SI-NEXT:    v_writelane_b32 v63, s80, 24
; SI-NEXT:    v_writelane_b32 v63, s81, 25
; SI-NEXT:    v_writelane_b32 v63, s82, 26
; SI-NEXT:    v_writelane_b32 v63, s83, 27
; SI-NEXT:    v_writelane_b32 v63, s84, 28
; SI-NEXT:    v_writelane_b32 v63, s85, 29
; SI-NEXT:    v_writelane_b32 v63, s86, 30
; SI-NEXT:    v_writelane_b32 v63, s87, 31
; SI-NEXT:    v_writelane_b32 v63, s96, 32
; SI-NEXT:    v_writelane_b32 v63, s97, 33
; SI-NEXT:    v_writelane_b32 v63, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v63, s99, 35
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s10, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s12, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s14, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s42, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s47, v0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB41_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s6, 0xffff0000
; SI-NEXT:    v_writelane_b32 v62, s4, 3
; SI-NEXT:    s_lshl_b32 s4, s6, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 2
; SI-NEXT:    s_and_b32 s4, s7, 0xffff0000
; SI-NEXT:    v_writelane_b32 v62, s4, 1
; SI-NEXT:    s_lshl_b32 s4, s7, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 0
; SI-NEXT:    s_and_b32 s60, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s61, s8, 16
; SI-NEXT:    s_and_b32 s62, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s63, s9, 16
; SI-NEXT:    s_and_b32 s72, s10, 0xffff0000
; SI-NEXT:    s_lshl_b32 s73, s10, 16
; SI-NEXT:    s_and_b32 s74, s11, 0xffff0000
; SI-NEXT:    s_lshl_b32 s75, s11, 16
; SI-NEXT:    s_and_b32 s76, s12, 0xffff0000
; SI-NEXT:    s_lshl_b32 s77, s12, 16
; SI-NEXT:    s_and_b32 s78, s13, 0xffff0000
; SI-NEXT:    s_lshl_b32 s79, s13, 16
; SI-NEXT:    s_and_b32 s88, s14, 0xffff0000
; SI-NEXT:    s_lshl_b32 s89, s14, 16
; SI-NEXT:    s_and_b32 s90, s15, 0xffff0000
; SI-NEXT:    s_lshl_b32 s91, s15, 16
; SI-NEXT:    s_and_b32 s92, s40, 0xffff0000
; SI-NEXT:    s_lshl_b32 s93, s40, 16
; SI-NEXT:    s_and_b32 s94, s41, 0xffff0000
; SI-NEXT:    s_lshl_b32 s95, s41, 16
; SI-NEXT:    s_and_b32 s30, s42, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, s42, 16
; SI-NEXT:    s_and_b32 s34, s43, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, s43, 16
; SI-NEXT:    s_and_b32 s36, s44, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, s44, 16
; SI-NEXT:    s_and_b32 s38, s45, 0xffff0000
; SI-NEXT:    s_lshl_b32 s39, s45, 16
; SI-NEXT:    s_and_b32 s48, s46, 0xffff0000
; SI-NEXT:    s_lshl_b32 s49, s46, 16
; SI-NEXT:    s_and_b32 s50, s47, 0xffff0000
; SI-NEXT:    s_lshl_b32 s51, s47, 16
; SI-NEXT:    s_and_b32 s52, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s53, s29, 16
; SI-NEXT:    s_and_b32 s54, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s55, s28, 16
; SI-NEXT:    s_and_b32 s64, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s65, s27, 16
; SI-NEXT:    s_and_b32 s66, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s67, s26, 16
; SI-NEXT:    s_and_b32 s68, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s69, s25, 16
; SI-NEXT:    s_and_b32 s70, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s71, s24, 16
; SI-NEXT:    s_and_b32 s80, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s81, s23, 16
; SI-NEXT:    s_and_b32 s82, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s83, s22, 16
; SI-NEXT:    s_and_b32 s84, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s85, s21, 16
; SI-NEXT:    s_and_b32 s86, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s87, s20, 16
; SI-NEXT:    s_and_b32 s96, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s97, s19, 16
; SI-NEXT:    s_and_b32 s98, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s99, s18, 16
; SI-NEXT:    s_and_b32 s56, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s57, s17, 16
; SI-NEXT:    s_and_b32 s58, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s59, s16, 16
; SI-NEXT:    s_cbranch_execnz .LBB41_4
; SI-NEXT:  .LBB41_2: ; %cmp.true
; SI-NEXT:    v_add_f32_e64 v29, s6, 1.0
; SI-NEXT:    v_add_f32_e64 v28, s7, 1.0
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_add_f32_e64 v27, s8, 1.0
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_add_f32_e64 v0, s18, 1.0
; SI-NEXT:    v_add_f32_e64 v4, s22, 1.0
; SI-NEXT:    v_add_f32_e64 v5, s23, 1.0
; SI-NEXT:    v_add_f32_e64 v12, s47, 1.0
; SI-NEXT:    v_add_f32_e64 v13, s46, 1.0
; SI-NEXT:    v_add_f32_e64 v14, s45, 1.0
; SI-NEXT:    v_add_f32_e64 v15, s44, 1.0
; SI-NEXT:    v_add_f32_e64 v26, s9, 1.0
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_add_f32_e64 v2, s20, 1.0
; SI-NEXT:    v_add_f32_e64 v3, s21, 1.0
; SI-NEXT:    v_add_f32_e64 v9, s27, 1.0
; SI-NEXT:    v_add_f32_e64 v10, s28, 1.0
; SI-NEXT:    v_add_f32_e64 v11, s29, 1.0
; SI-NEXT:    v_add_f32_e64 v22, s13, 1.0
; SI-NEXT:    v_add_f32_e64 v23, s12, 1.0
; SI-NEXT:    v_add_f32_e64 v24, s11, 1.0
; SI-NEXT:    v_add_f32_e64 v25, s10, 1.0
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v15
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v13
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v13
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v12
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v5
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v0
; SI-NEXT:    v_add_f32_e64 v0, s17, 1.0
; SI-NEXT:    v_add_f32_e64 v1, s19, 1.0
; SI-NEXT:    v_add_f32_e64 v6, s24, 1.0
; SI-NEXT:    v_add_f32_e64 v7, s25, 1.0
; SI-NEXT:    v_add_f32_e64 v8, s26, 1.0
; SI-NEXT:    v_add_f32_e64 v16, s43, 1.0
; SI-NEXT:    v_add_f32_e64 v17, s42, 1.0
; SI-NEXT:    v_add_f32_e64 v18, s41, 1.0
; SI-NEXT:    v_add_f32_e64 v19, s40, 1.0
; SI-NEXT:    v_add_f32_e64 v20, s15, 1.0
; SI-NEXT:    v_add_f32_e64 v21, s14, 1.0
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v25
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v24
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v23
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v22
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v11
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v11
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v10
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v3
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v0
; SI-NEXT:    v_add_f32_e64 v0, s16, 1.0
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v21
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v20
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v19
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v18
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v17
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v16
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v8
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v7
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v6
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_branch .LBB41_5
; SI-NEXT:  .LBB41_3:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr59
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr57
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr99
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr97
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr87
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr85
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr83
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr81
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr71
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr37
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr35
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr31
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr93
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr77
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr75
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr73
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr63
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr61
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    s_branch .LBB41_2
; SI-NEXT:  .LBB41_4:
; SI-NEXT:    v_mov_b32_e32 v7, s63
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v7, s61
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v7, s60
; SI-NEXT:    v_readlane_b32 s4, v62, 0
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v7, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 1
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v7, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 2
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v7, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 3
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v7, s4
; SI-NEXT:    v_mov_b32_e32 v0, s59
; SI-NEXT:    v_mov_b32_e32 v1, s58
; SI-NEXT:    v_mov_b32_e32 v2, s57
; SI-NEXT:    v_mov_b32_e32 v3, s56
; SI-NEXT:    v_mov_b32_e32 v4, s99
; SI-NEXT:    v_mov_b32_e32 v5, s98
; SI-NEXT:    v_mov_b32_e32 v6, s97
; SI-NEXT:    v_mov_b32_e32 v39, s96
; SI-NEXT:    v_mov_b32_e32 v38, s87
; SI-NEXT:    v_mov_b32_e32 v9, s86
; SI-NEXT:    v_mov_b32_e32 v10, s85
; SI-NEXT:    v_mov_b32_e32 v11, s84
; SI-NEXT:    v_mov_b32_e32 v12, s83
; SI-NEXT:    v_mov_b32_e32 v13, s82
; SI-NEXT:    v_mov_b32_e32 v14, s81
; SI-NEXT:    v_mov_b32_e32 v15, s80
; SI-NEXT:    v_mov_b32_e32 v48, s71
; SI-NEXT:    v_mov_b32_e32 v17, s70
; SI-NEXT:    v_mov_b32_e32 v18, s69
; SI-NEXT:    v_mov_b32_e32 v19, s68
; SI-NEXT:    v_mov_b32_e32 v20, s67
; SI-NEXT:    v_mov_b32_e32 v21, s66
; SI-NEXT:    v_mov_b32_e32 v22, s65
; SI-NEXT:    v_mov_b32_e32 v23, s64
; SI-NEXT:    v_mov_b32_e32 v24, s55
; SI-NEXT:    v_mov_b32_e32 v25, s54
; SI-NEXT:    v_mov_b32_e32 v26, s53
; SI-NEXT:    v_mov_b32_e32 v27, s52
; SI-NEXT:    v_mov_b32_e32 v28, s51
; SI-NEXT:    v_mov_b32_e32 v29, s50
; SI-NEXT:    v_mov_b32_e32 v30, s49
; SI-NEXT:    v_mov_b32_e32 v31, s48
; SI-NEXT:    v_mov_b32_e32 v32, s39
; SI-NEXT:    v_mov_b32_e32 v33, s38
; SI-NEXT:    v_mov_b32_e32 v36, s37
; SI-NEXT:    v_mov_b32_e32 v37, s36
; SI-NEXT:    v_mov_b32_e32 v60, s35
; SI-NEXT:    v_mov_b32_e32 v61, s34
; SI-NEXT:    v_mov_b32_e32 v58, s31
; SI-NEXT:    v_mov_b32_e32 v59, s30
; SI-NEXT:    v_mov_b32_e32 v56, s95
; SI-NEXT:    v_mov_b32_e32 v57, s94
; SI-NEXT:    v_mov_b32_e32 v46, s93
; SI-NEXT:    v_mov_b32_e32 v47, s92
; SI-NEXT:    v_mov_b32_e32 v44, s91
; SI-NEXT:    v_mov_b32_e32 v45, s90
; SI-NEXT:    v_mov_b32_e32 v42, s89
; SI-NEXT:    v_mov_b32_e32 v43, s88
; SI-NEXT:    v_mov_b32_e32 v40, s79
; SI-NEXT:    v_mov_b32_e32 v41, s78
; SI-NEXT:    v_mov_b32_e32 v54, s77
; SI-NEXT:    v_mov_b32_e32 v55, s76
; SI-NEXT:    v_mov_b32_e32 v52, s75
; SI-NEXT:    v_mov_b32_e32 v53, s74
; SI-NEXT:    v_mov_b32_e32 v50, s73
; SI-NEXT:    v_mov_b32_e32 v51, s72
; SI-NEXT:    v_mov_b32_e32 v49, s62
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:  .LBB41_5: ; %end
; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v39
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v9
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v13
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v37
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v4
; SI-NEXT:    v_lshr_b64 v[33:34], v[34:35], 16
; SI-NEXT:    v_lshr_b64 v[34:35], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[2:3], v[7:8], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v38
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v36
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v60
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v58
; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v46
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v44
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v42
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v40
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v50
; SI-NEXT:    v_readlane_b32 s99, v63, 35
; SI-NEXT:    v_readlane_b32 s98, v63, 34
; SI-NEXT:    v_readlane_b32 s97, v63, 33
; SI-NEXT:    v_readlane_b32 s96, v63, 32
; SI-NEXT:    v_readlane_b32 s87, v63, 31
; SI-NEXT:    v_readlane_b32 s86, v63, 30
; SI-NEXT:    v_readlane_b32 s85, v63, 29
; SI-NEXT:    v_readlane_b32 s84, v63, 28
; SI-NEXT:    v_readlane_b32 s83, v63, 27
; SI-NEXT:    v_readlane_b32 s82, v63, 26
; SI-NEXT:    v_readlane_b32 s81, v63, 25
; SI-NEXT:    v_readlane_b32 s80, v63, 24
; SI-NEXT:    v_readlane_b32 s71, v63, 23
; SI-NEXT:    v_readlane_b32 s70, v63, 22
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    v_lshr_b64 v[4:5], v[5:6], 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[7:8], v[15:16], 16
; SI-NEXT:    v_lshr_b64 v[8:9], v[9:10], 16
; SI-NEXT:    v_lshr_b64 v[9:10], v[11:12], 16
; SI-NEXT:    v_lshr_b64 v[10:11], v[13:14], 16
; SI-NEXT:    v_lshr_b64 v[11:12], v[22:23], 16
; SI-NEXT:    v_lshr_b64 v[12:13], v[24:25], 16
; SI-NEXT:    v_lshr_b64 v[13:14], v[26:27], 16
; SI-NEXT:    v_lshr_b64 v[14:15], v[28:29], 16
; SI-NEXT:    v_lshr_b64 v[15:16], v[30:31], 16
; SI-NEXT:    v_lshr_b64 v[16:17], v[17:18], 16
; SI-NEXT:    v_lshr_b64 v[17:18], v[19:20], 16
; SI-NEXT:    v_lshr_b64 v[18:19], v[60:61], 16
; SI-NEXT:    v_lshr_b64 v[19:20], v[58:59], 16
; SI-NEXT:    v_lshr_b64 v[20:21], v[56:57], 16
; SI-NEXT:    v_lshr_b64 v[21:22], v[46:47], 16
; SI-NEXT:    v_lshr_b64 v[22:23], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[23:24], v[42:43], 16
; SI-NEXT:    v_lshr_b64 v[24:25], v[40:41], 16
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[25:26], v[54:55], 16
; SI-NEXT:    v_lshr_b64 v[26:27], v[52:53], 16
; SI-NEXT:    v_lshr_b64 v[27:28], v[50:51], 16
; SI-NEXT:    v_lshr_b64 v[28:29], v[48:49], 16
; SI-NEXT:    v_lshr_b64 v[29:30], v[38:39], 16
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[30:31], v[36:37], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshr_b64 v[31:32], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v0, v33
; SI-NEXT:    v_mov_b32_e32 v1, v34
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v64bf16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s36, 2
; VI-NEXT:    v_writelane_b32 v32, s37, 3
; VI-NEXT:    v_writelane_b32 v32, s38, 4
; VI-NEXT:    v_writelane_b32 v32, s39, 5
; VI-NEXT:    v_writelane_b32 v32, s48, 6
; VI-NEXT:    v_writelane_b32 v32, s49, 7
; VI-NEXT:    v_writelane_b32 v32, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB41_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB41_4
; VI-NEXT:  .LBB41_2: ; %cmp.true
; VI-NEXT:    v_add_f32_e64 v15, s31, 1.0
; VI-NEXT:    v_add_f32_e64 v14, s30, 1.0
; VI-NEXT:    v_add_f32_e64 v13, s29, 1.0
; VI-NEXT:    v_add_f32_e64 v12, s28, 1.0
; VI-NEXT:    v_add_f32_e64 v11, s27, 1.0
; VI-NEXT:    v_add_f32_e64 v10, s26, 1.0
; VI-NEXT:    v_add_f32_e64 v9, s25, 1.0
; VI-NEXT:    v_add_f32_e64 v8, s24, 1.0
; VI-NEXT:    v_add_f32_e64 v7, s23, 1.0
; VI-NEXT:    v_add_f32_e64 v6, s22, 1.0
; VI-NEXT:    v_add_f32_e64 v5, s21, 1.0
; VI-NEXT:    v_add_f32_e64 v4, s20, 1.0
; VI-NEXT:    v_add_f32_e64 v3, s19, 1.0
; VI-NEXT:    v_add_f32_e64 v2, s18, 1.0
; VI-NEXT:    v_add_f32_e64 v1, s17, 1.0
; VI-NEXT:    v_add_f32_e64 v0, s16, 1.0
; VI-NEXT:    v_add_f32_e64 v31, s51, 1.0
; VI-NEXT:    v_add_f32_e64 v30, s50, 1.0
; VI-NEXT:    v_add_f32_e64 v29, s49, 1.0
; VI-NEXT:    v_add_f32_e64 v28, s48, 1.0
; VI-NEXT:    v_add_f32_e64 v27, s47, 1.0
; VI-NEXT:    v_add_f32_e64 v26, s46, 1.0
; VI-NEXT:    v_add_f32_e64 v25, s45, 1.0
; VI-NEXT:    v_add_f32_e64 v24, s44, 1.0
; VI-NEXT:    v_add_f32_e64 v23, s43, 1.0
; VI-NEXT:    v_add_f32_e64 v22, s42, 1.0
; VI-NEXT:    v_add_f32_e64 v21, s41, 1.0
; VI-NEXT:    v_add_f32_e64 v20, s40, 1.0
; VI-NEXT:    v_add_f32_e64 v19, s39, 1.0
; VI-NEXT:    v_add_f32_e64 v18, s38, 1.0
; VI-NEXT:    v_add_f32_e64 v17, s37, 1.0
; VI-NEXT:    v_add_f32_e64 v16, s36, 1.0
; VI-NEXT:    s_branch .LBB41_5
; VI-NEXT:  .LBB41_3:
; VI-NEXT:    s_branch .LBB41_2
; VI-NEXT:  .LBB41_4:
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:  .LBB41_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v32, 9
; VI-NEXT:    v_readlane_b32 s50, v32, 8
; VI-NEXT:    v_readlane_b32 s49, v32, 7
; VI-NEXT:    v_readlane_b32 s48, v32, 6
; VI-NEXT:    v_readlane_b32 s39, v32, 5
; VI-NEXT:    v_readlane_b32 s38, v32, 4
; VI-NEXT:    v_readlane_b32 s37, v32, 3
; VI-NEXT:    v_readlane_b32 s36, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v64bf16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB41_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB41_4
; GFX9-NEXT:  .LBB41_2: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e64 v15, s31, 1.0
; GFX9-NEXT:    v_add_f32_e64 v14, s30, 1.0
; GFX9-NEXT:    v_add_f32_e64 v13, s29, 1.0
; GFX9-NEXT:    v_add_f32_e64 v12, s28, 1.0
; GFX9-NEXT:    v_add_f32_e64 v11, s27, 1.0
; GFX9-NEXT:    v_add_f32_e64 v10, s26, 1.0
; GFX9-NEXT:    v_add_f32_e64 v9, s25, 1.0
; GFX9-NEXT:    v_add_f32_e64 v8, s24, 1.0
; GFX9-NEXT:    v_add_f32_e64 v7, s23, 1.0
; GFX9-NEXT:    v_add_f32_e64 v6, s22, 1.0
; GFX9-NEXT:    v_add_f32_e64 v5, s21, 1.0
; GFX9-NEXT:    v_add_f32_e64 v4, s20, 1.0
; GFX9-NEXT:    v_add_f32_e64 v3, s19, 1.0
; GFX9-NEXT:    v_add_f32_e64 v2, s18, 1.0
; GFX9-NEXT:    v_add_f32_e64 v1, s17, 1.0
; GFX9-NEXT:    v_add_f32_e64 v0, s16, 1.0
; GFX9-NEXT:    v_add_f32_e64 v31, s51, 1.0
; GFX9-NEXT:    v_add_f32_e64 v30, s50, 1.0
; GFX9-NEXT:    v_add_f32_e64 v29, s49, 1.0
; GFX9-NEXT:    v_add_f32_e64 v28, s48, 1.0
; GFX9-NEXT:    v_add_f32_e64 v27, s47, 1.0
; GFX9-NEXT:    v_add_f32_e64 v26, s46, 1.0
; GFX9-NEXT:    v_add_f32_e64 v25, s45, 1.0
; GFX9-NEXT:    v_add_f32_e64 v24, s44, 1.0
; GFX9-NEXT:    v_add_f32_e64 v23, s43, 1.0
; GFX9-NEXT:    v_add_f32_e64 v22, s42, 1.0
; GFX9-NEXT:    v_add_f32_e64 v21, s41, 1.0
; GFX9-NEXT:    v_add_f32_e64 v20, s40, 1.0
; GFX9-NEXT:    v_add_f32_e64 v19, s39, 1.0
; GFX9-NEXT:    v_add_f32_e64 v18, s38, 1.0
; GFX9-NEXT:    v_add_f32_e64 v17, s37, 1.0
; GFX9-NEXT:    v_add_f32_e64 v16, s36, 1.0
; GFX9-NEXT:    s_branch .LBB41_5
; GFX9-NEXT:  .LBB41_3:
; GFX9-NEXT:    s_branch .LBB41_2
; GFX9-NEXT:  .LBB41_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:  .LBB41_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v64bf16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB41_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB41_4
; GFX11-NEXT:  .LBB41_2: ; %cmp.true
; GFX11-NEXT:    v_add_f32_e64 v15, s27, 1.0
; GFX11-NEXT:    v_add_f32_e64 v14, s26, 1.0
; GFX11-NEXT:    v_add_f32_e64 v13, s25, 1.0
; GFX11-NEXT:    v_add_f32_e64 v12, s24, 1.0
; GFX11-NEXT:    v_add_f32_e64 v11, s23, 1.0
; GFX11-NEXT:    v_add_f32_e64 v10, s22, 1.0
; GFX11-NEXT:    v_add_f32_e64 v9, s21, 1.0
; GFX11-NEXT:    v_add_f32_e64 v8, s20, 1.0
; GFX11-NEXT:    v_add_f32_e64 v7, s19, 1.0
; GFX11-NEXT:    v_add_f32_e64 v6, s18, 1.0
; GFX11-NEXT:    v_add_f32_e64 v5, s17, 1.0
; GFX11-NEXT:    v_add_f32_e64 v4, s16, 1.0
; GFX11-NEXT:    v_add_f32_e64 v3, s15, 1.0
; GFX11-NEXT:    v_add_f32_e64 v2, s14, 1.0
; GFX11-NEXT:    v_add_f32_e64 v1, s13, 1.0
; GFX11-NEXT:    v_add_f32_e64 v0, s12, 1.0
; GFX11-NEXT:    v_add_f32_e64 v31, s51, 1.0
; GFX11-NEXT:    v_add_f32_e64 v30, s50, 1.0
; GFX11-NEXT:    v_add_f32_e64 v29, s49, 1.0
; GFX11-NEXT:    v_add_f32_e64 v28, s48, 1.0
; GFX11-NEXT:    v_add_f32_e64 v27, s47, 1.0
; GFX11-NEXT:    v_add_f32_e64 v26, s46, 1.0
; GFX11-NEXT:    v_add_f32_e64 v25, s45, 1.0
; GFX11-NEXT:    v_add_f32_e64 v24, s44, 1.0
; GFX11-NEXT:    v_add_f32_e64 v23, s43, 1.0
; GFX11-NEXT:    v_add_f32_e64 v22, s42, 1.0
; GFX11-NEXT:    v_add_f32_e64 v21, s41, 1.0
; GFX11-NEXT:    v_add_f32_e64 v20, s40, 1.0
; GFX11-NEXT:    v_add_f32_e64 v19, s39, 1.0
; GFX11-NEXT:    v_add_f32_e64 v18, s38, 1.0
; GFX11-NEXT:    v_add_f32_e64 v17, s37, 1.0
; GFX11-NEXT:    v_add_f32_e64 v16, s36, 1.0
; GFX11-NEXT:    s_branch .LBB41_5
; GFX11-NEXT:  .LBB41_3:
; GFX11-NEXT:    s_branch .LBB41_2
; GFX11-NEXT:  .LBB41_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:  .LBB41_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define <32 x float> @bitcast_v64bf16_to_v32f32(<64 x bfloat> %a, i32 %b) {
; SI-LABEL: bitcast_v64bf16_to_v32f32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v0
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v63
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v62
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v60
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v58, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_and_b32_e32 v46, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v36, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v53
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v37
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v38
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v38
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v42
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v40
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v55
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v18
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v19
; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v51
; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v20
; SI-NEXT:    v_mul_f32_e32 v45, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v49
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v41, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v23
; SI-NEXT:    v_mul_f32_e32 v55, 1.0, v39
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v53, 1.0, v36
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v25
; SI-NEXT:    v_mul_f32_e32 v51, 1.0, v35
; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v49, 1.0, v34
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v39, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v29
; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v31
; SI-NEXT:    v_mul_f32_e32 v35, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v33, 1.0, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB42_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v51
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v39
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v36
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v34
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_alignbit_b32 v19, v19, v59, 16
; SI-NEXT:    v_alignbit_b32 v20, v20, v57, 16
; SI-NEXT:    v_alignbit_b32 v21, v21, v46, 16
; SI-NEXT:    v_alignbit_b32 v22, v22, v44, 16
; SI-NEXT:    v_alignbit_b32 v23, v23, v42, 16
; SI-NEXT:    v_alignbit_b32 v24, v24, v40, 16
; SI-NEXT:    v_alignbit_b32 v25, v25, v54, 16
; SI-NEXT:    v_alignbit_b32 v26, v26, v52, 16
; SI-NEXT:    v_alignbit_b32 v27, v27, v50, 16
; SI-NEXT:    v_alignbit_b32 v28, v28, v48, 16
; SI-NEXT:    v_alignbit_b32 v29, v29, v38, 16
; SI-NEXT:    v_alignbit_b32 v30, v30, v35, 16
; SI-NEXT:    v_alignbit_b32 v31, v31, v33, 16
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_alignbit_b32 v0, v0, v63, 16
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v62
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v60
; SI-NEXT:    v_alignbit_b32 v18, v18, v61, 16
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:  .LBB42_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB42_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v63
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v60
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v55
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v0, v1, v0, 16
; SI-NEXT:    v_alignbit_b32 v1, v3, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v2, v3, v2, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_alignbit_b32 v3, v4, v3, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_alignbit_b32 v4, v5, v4, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_alignbit_b32 v5, v6, v5, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_alignbit_b32 v6, v7, v6, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_alignbit_b32 v7, v8, v7, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_alignbit_b32 v8, v9, v8, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_alignbit_b32 v9, v10, v9, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_alignbit_b32 v10, v11, v10, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_alignbit_b32 v11, v12, v11, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_alignbit_b32 v12, v13, v12, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v13, v14, v13, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_alignbit_b32 v14, v15, v14, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v15, v16, v15, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v16, v17, v16, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_alignbit_b32 v17, v18, v17, 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v61
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_alignbit_b32 v18, v19, v18, 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v59
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_alignbit_b32 v19, v20, v19, 16
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v57
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_alignbit_b32 v20, v21, v20, 16
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_alignbit_b32 v21, v22, v21, 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v44
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_alignbit_b32 v22, v23, v22, 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_alignbit_b32 v23, v24, v23, 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_alignbit_b32 v24, v25, v24, 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_alignbit_b32 v25, v26, v25, 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_alignbit_b32 v26, v27, v26, 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_alignbit_b32 v27, v28, v27, 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_alignbit_b32 v28, v29, v28, 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_alignbit_b32 v29, v30, v29, 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v35
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_alignbit_b32 v30, v31, v30, 16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v33
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_alignbit_b32 v31, v32, v31, 16
; SI-NEXT:  .LBB42_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v32f32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB42_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v15, 16, 1
; VI-NEXT:    s_movk_i32 s6, 0x7fff
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v15
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_alignbit_b32 v15, v15, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v14, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v14
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; VI-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_alignbit_b32 v14, v14, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v13
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_alignbit_b32 v13, v13, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v12, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v12
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; VI-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_alignbit_b32 v12, v12, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v11
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_alignbit_b32 v11, v11, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v10, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v10
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; VI-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_alignbit_b32 v10, v10, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v9
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_alignbit_b32 v9, v9, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v8, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v8
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; VI-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_alignbit_b32 v8, v8, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v7
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_alignbit_b32 v7, v7, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v6, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v6
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; VI-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_alignbit_b32 v6, v6, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v5
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_alignbit_b32 v5, v5, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v4, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v4
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; VI-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_alignbit_b32 v4, v4, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v3
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_alignbit_b32 v3, v3, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v2
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_alignbit_b32 v2, v2, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v1
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_alignbit_b32 v1, v1, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v0
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; VI-NEXT:    v_alignbit_b32 v0, v0, v32, 16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v30, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v30
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; VI-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_alignbit_b32 v30, v30, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v29, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v29
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_alignbit_b32 v29, v29, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v28, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v28
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; VI-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_alignbit_b32 v28, v28, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v27
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_alignbit_b32 v27, v27, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v26, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v26
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; VI-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_alignbit_b32 v26, v26, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v25
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_alignbit_b32 v25, v25, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v24, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v24
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; VI-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_alignbit_b32 v24, v24, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v23
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_alignbit_b32 v23, v23, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v22, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v22
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; VI-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_alignbit_b32 v22, v22, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v21
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_alignbit_b32 v21, v21, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v20, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v20
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; VI-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_alignbit_b32 v20, v20, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v19
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_alignbit_b32 v19, v19, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v18
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_alignbit_b32 v18, v18, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v17
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_alignbit_b32 v17, v17, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v16
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_alignbit_b32 v16, v16, v32, 16
; VI-NEXT:  .LBB42_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v32f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB42_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    s_movk_i32 s6, 0x7fff
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v15, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v15, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; GFX9-NEXT:    s_mov_b32 s7, 0x7060302
; GFX9-NEXT:    v_perm_b32 v15, v15, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v14, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v14, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v14, v14, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v13, v13, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v12, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v12, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v12, v12, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v11, v11, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v10, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v10, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v10, v10, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v9, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v9, v9, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v8, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v8, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v8, v8, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v7, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v7, v7, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v6, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v6, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v6, v6, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v5, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v5, v5, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v4, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v4, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v4, v4, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v3, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v3, v3, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v2, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v2, v2, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v1, v1, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v0, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v0, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v0, v0, v32, s7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v30, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v30, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v30, v30, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v29, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v29, v29, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v28, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v28, v28, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v27, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v27, v27, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v26, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v26, v26, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v25, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v25, v25, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v24, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v24, v24, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v23, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v23, v23, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v22, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v22, v22, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v21, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v21, v21, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v20, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v20, v20, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v19, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v19, v19, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v18, v18, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v17, v17, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v16, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v16, v16, v32, s7
; GFX9-NEXT:  .LBB42_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v32f32:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB42_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v63, v31 :: v_dual_mov_b32 v62, v30
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v61, v29 :: v_dual_mov_b32 v60, v28
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v59, v27 :: v_dual_mov_b32 v58, v26
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v57, v25 :: v_dual_mov_b32 v56, v24
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v23 :: v_dual_mov_b32 v54, v22
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v21 :: v_dual_mov_b32 v52, v20
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v19 :: v_dual_mov_b32 v50, v18
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v17 :: v_dual_mov_b32 v48, v16
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v47, v15 :: v_dual_mov_b32 v46, v14
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v45, v13 :: v_dual_mov_b32 v44, v12
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v43, v11 :: v_dual_mov_b32 v42, v10
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v41, v9 :: v_dual_mov_b32 v40, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v7 :: v_dual_mov_b32 v38, v6
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v5 :: v_dual_mov_b32 v36, v4
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v3 :: v_dual_mov_b32 v34, v2
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v1 :: v_dual_mov_b32 v32, v0
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT:  .LBB42_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB42_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v32, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v35, v38 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v39, v33, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v47, v36, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v32.h
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_add3_u32 v38, v48, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v38, v49, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v34 :: v_dual_cndmask_b32 v46, v35, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, v14.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v33, v34, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v13, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v45, v33, v35
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v13.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v15, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v44, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, v12.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v11, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v14, v32 :: v_dual_and_b32 v12, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v43, v15, v33
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v10, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v13, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v42, v14, v32, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v10.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v9, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v41, v13, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, v9.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v8, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v11, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v40, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v8.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v7, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v39, v11, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v7.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v9, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v6.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v8, v10 :: v_dual_and_b32 v6, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v37, v9, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v8, v10, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v4.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v6, v8 :: v_dual_and_b32 v4, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_cndmask_b32 v35, v7, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v5, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v6, v8, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v4, v6 :: v_dual_and_b32 v2, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_cndmask_b32 v33, v5, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v1.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v31
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v5, 16, v31
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v3, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v4, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v0.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v63, v3, v7 :: v_dual_and_b32 v0, 0xffff0000, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v62, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v28
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v61, v3, v7 :: v_dual_lshlrev_b32 v4, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v60, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v59, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v25
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v4, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v58, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v24
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v57, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v56, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v55, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v5, 16, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v53, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v52, v0, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v2.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_lshlrev_b32 v6, 16, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v6 :: v_dual_cndmask_b32 v51, v0, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v8, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v50, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v5, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v16
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v1 :: v_dual_add_f32 v1, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v0.h
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v9, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v1, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v5.h
; GFX11-TRUE16-NEXT:  .LBB42_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v40 :: v_dual_mov_b32 v9, v41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v42 :: v_dual_mov_b32 v11, v43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, v44 :: v_dual_mov_b32 v13, v45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v46 :: v_dual_mov_b32 v15, v47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v56 :: v_dual_mov_b32 v25, v57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v58 :: v_dual_mov_b32 v27, v59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v60 :: v_dual_mov_b32 v29, v61
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v62 :: v_dual_mov_b32 v31, v63
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v33
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v34 :: v_dual_mov_b32 v3, v35
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v36 :: v_dual_mov_b32 v5, v37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v38 :: v_dual_mov_b32 v7, v39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v48 :: v_dual_mov_b32 v17, v49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v50 :: v_dual_mov_b32 v19, v51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v52 :: v_dual_mov_b32 v21, v53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v54 :: v_dual_mov_b32 v23, v55
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v32f32:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB42_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v39, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v13, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v11, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v8, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v9, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v34 :: v_dual_add_f32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v6, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v6, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v7, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v5, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v3, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v0, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v1, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v31
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v31, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v29
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v30, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v29, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v28, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v27
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v28, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v27, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v26, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v25, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v23
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v23, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v36, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v21, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v32, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v38, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v20, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v19, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v48, v18, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v36, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v36
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v39, v17, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v16, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v35, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v37, v50, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v16, v36, 0x7060302
; GFX11-FAKE16-NEXT:  .LBB42_2: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define inreg <32 x float> @bitcast_v64bf16_to_v32f32_scalar(<64 x bfloat> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64bf16_to_v32f32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    s_and_b32 s12, s25, 0xffff0000
; SI-NEXT:    s_and_b32 s30, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v1
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; SI-NEXT:    s_and_b32 s34, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_readfirstlane_b32 s94, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s34
; SI-NEXT:    v_readfirstlane_b32 s90, v5
; SI-NEXT:    s_and_b32 s95, s94, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s30
; SI-NEXT:    v_readfirstlane_b32 s88, v6
; SI-NEXT:    s_and_b32 s91, s90, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s95
; SI-NEXT:    v_readfirstlane_b32 s78, v7
; SI-NEXT:    s_and_b32 s89, s88, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s91
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    s_and_b32 s79, s78, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s89
; SI-NEXT:    v_readfirstlane_b32 s74, v9
; SI-NEXT:    s_and_b32 s77, s76, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s79
; SI-NEXT:    v_readfirstlane_b32 s72, v10
; SI-NEXT:    s_and_b32 s75, s74, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s77
; SI-NEXT:    v_readfirstlane_b32 s62, v11
; SI-NEXT:    s_and_b32 s73, s72, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s75
; SI-NEXT:    v_readfirstlane_b32 s60, v12
; SI-NEXT:    s_and_b32 s63, s62, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s73
; SI-NEXT:    v_readfirstlane_b32 s58, v13
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s63
; SI-NEXT:    v_readfirstlane_b32 s56, v14
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s61
; SI-NEXT:    v_readfirstlane_b32 s46, v15
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s59
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s57
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s47
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    v_readfirstlane_b32 s92, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s45
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s29, 16
; SI-NEXT:    s_lshl_b32 s7, s28, 16
; SI-NEXT:    s_and_b32 s8, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s9, s27, 16
; SI-NEXT:    s_and_b32 s10, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s11, s26, 16
; SI-NEXT:    s_lshl_b32 s13, s25, 16
; SI-NEXT:    s_and_b32 s14, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s15, s24, 16
; SI-NEXT:    s_and_b32 s24, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s25, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s26, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s27, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s28, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s29, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_and_b32 s41, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_and_b32 s93, s92, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_and_b32 s36, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s43
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cmp_lg_u32 vcc_lo, 0
; SI-NEXT:    v_mul_f32_e64 v55, 1.0, s41
; SI-NEXT:    v_mul_f32_e64 v53, 1.0, s40
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s29
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v62, 1.0, s27
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s26
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s25
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s24
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s14
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s10
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s8
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s4
; SI-NEXT:    v_mul_f32_e64 v39, 1.0, s36
; SI-NEXT:    v_mul_f32_e64 v47, 1.0, s93
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e64 v43, 1.0, s16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s17
; SI-NEXT:    v_mul_f32_e64 v34, 1.0, s18
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s19
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s20
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s21
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s22
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s15
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s13
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s11
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s9
; SI-NEXT:    v_mul_f32_e64 v61, 1.0, s7
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s5
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s37
; SI-NEXT:    v_mul_f32_e64 v60, 1.0, s35
; SI-NEXT:    v_mul_f32_e64 v33, 1.0, s31
; SI-NEXT:    v_mul_f32_e64 v57, 1.0, s94
; SI-NEXT:    v_mul_f32_e64 v59, 1.0, s92
; SI-NEXT:    v_mul_f32_e64 v56, 1.0, s90
; SI-NEXT:    v_mul_f32_e64 v38, 1.0, s88
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s78
; SI-NEXT:    v_mul_f32_e64 v36, 1.0, s76
; SI-NEXT:    v_mul_f32_e64 v46, 1.0, s74
; SI-NEXT:    v_mul_f32_e64 v48, 1.0, s72
; SI-NEXT:    v_mul_f32_e64 v49, 1.0, s62
; SI-NEXT:    v_mul_f32_e64 v45, 1.0, s60
; SI-NEXT:    v_mul_f32_e64 v42, 1.0, s58
; SI-NEXT:    v_mul_f32_e64 v51, 1.0, s56
; SI-NEXT:    v_mul_f32_e64 v41, 1.0, s46
; SI-NEXT:    v_mul_f32_e64 v54, 1.0, s44
; SI-NEXT:    v_mul_f32_e64 v52, 1.0, s42
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB43_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v55
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v53
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v44, v1
; SI-NEXT:    v_mov_b32_e32 v43, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v12
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    v_mov_b32_e32 v36, v47
; SI-NEXT:    v_mov_b32_e32 v50, v49
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v40, v41
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v1, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[43:44], 16
; SI-NEXT:    v_mov_b32_e32 v2, v34
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[2:3], v[34:35], 16
; SI-NEXT:    v_mov_b32_e32 v3, v17
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[3:4], v[17:18], 16
; SI-NEXT:    v_mov_b32_e32 v4, v20
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[4:5], v[20:21], 16
; SI-NEXT:    v_mov_b32_e32 v5, v19
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v6
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[5:6], v[19:20], 16
; SI-NEXT:    v_mov_b32_e32 v6, v13
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[6:7], v[13:14], 16
; SI-NEXT:    v_mov_b32_e32 v7, v31
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[7:8], v[31:32], 16
; SI-NEXT:    v_mov_b32_e32 v8, v23
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[8:9], v[23:24], 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v11
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v44, v45
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v9
; SI-NEXT:    v_mov_b32_e32 v9, v25
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[9:10], v[25:26], 16
; SI-NEXT:    v_mov_b32_e32 v10, v30
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v11, v27
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[11:12], v[27:28], 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v18
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v12
; SI-NEXT:    v_mov_b32_e32 v12, v61
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[12:13], v[61:62], 16
; SI-NEXT:    v_mov_b32_e32 v13, v29
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[13:14], v[29:30], 16
; SI-NEXT:    v_mov_b32_e32 v14, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v62, v39
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[14:15], v[15:16], 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v15
; SI-NEXT:    v_mov_b32_e32 v15, v60
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[15:16], v[60:61], 16
; SI-NEXT:    v_mov_b32_e32 v16, v33
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[16:17], v[33:34], 16
; SI-NEXT:    v_mov_b32_e32 v17, v57
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[17:18], v[57:58], 16
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v47
; SI-NEXT:    v_lshr_b64 v[18:19], v[59:60], 16
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v20
; SI-NEXT:    v_lshr_b64 v[19:20], v[56:57], 16
; SI-NEXT:    v_mov_b32_e32 v20, v38
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[20:21], v[38:39], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; SI-NEXT:    v_mov_b32_e32 v39, v62
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v21
; SI-NEXT:    v_mov_b32_e32 v21, v22
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[21:22], v[22:23], 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v55
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[37:38], 16
; SI-NEXT:    v_lshr_b64 v[23:24], v[46:47], 16
; SI-NEXT:    v_lshr_b64 v[24:25], v[48:49], 16
; SI-NEXT:    v_lshr_b64 v[25:26], v[50:51], 16
; SI-NEXT:    v_lshr_b64 v[26:27], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[27:28], v[42:43], 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_lshr_b64 v[61:62], v[52:53], 16
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v45, v44
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v47, v36
; SI-NEXT:    v_mov_b32_e32 v36, v37
; SI-NEXT:    v_mov_b32_e32 v49, v50
; SI-NEXT:    v_mov_b32_e32 v51, v33
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_lshr_b64 v[29:30], v[40:41], 16
; SI-NEXT:    v_lshr_b64 v[30:31], v[54:55], 16
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v31, v61
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    s_branch .LBB43_3
; SI-NEXT:  .LBB43_2:
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB43_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v37, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v33, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(3)
; SI-NEXT:    v_mov_b32_e32 v38, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v50, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v40, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_cbranch_vccnz .LBB43_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v55
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v34, v56
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshr_b64 v[1:2], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v33
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_lshr_b64 v[32:33], v[32:33], 16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v37
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v31, v32
; SI-NEXT:  .LBB43_5: ; %end
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v32f32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v35, s36, 0
; VI-NEXT:    v_writelane_b32 v35, s37, 1
; VI-NEXT:    v_writelane_b32 v35, s38, 2
; VI-NEXT:    v_writelane_b32 v35, s39, 3
; VI-NEXT:    v_writelane_b32 v35, s48, 4
; VI-NEXT:    v_writelane_b32 v35, s49, 5
; VI-NEXT:    v_writelane_b32 v35, s50, 6
; VI-NEXT:    v_writelane_b32 v35, s51, 7
; VI-NEXT:    v_writelane_b32 v35, s52, 8
; VI-NEXT:    v_writelane_b32 v35, s53, 9
; VI-NEXT:    v_writelane_b32 v35, s54, 10
; VI-NEXT:    v_writelane_b32 v35, s55, 11
; VI-NEXT:    v_writelane_b32 v35, s64, 12
; VI-NEXT:    v_writelane_b32 v35, s65, 13
; VI-NEXT:    v_writelane_b32 v35, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v35, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB43_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB43_4
; VI-NEXT:  .LBB43_2: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x40c00000
; VI-NEXT:    s_lshl_b32 s4, s50, 16
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    s_and_b32 s4, s50, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v2, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s48, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; VI-NEXT:    v_add_f32_e32 v2, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    s_and_b32 s4, s48, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v4, v3, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_lshrrev_b64 v[14:15], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s46, 16
; VI-NEXT:    v_lshrrev_b64 v[12:13], 16, v[2:3]
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[10:11], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[8:9], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s38, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[4:5], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s36, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[2:3], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; VI-NEXT:    v_bfe_u32 v3, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v0
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_lshrrev_b64 v[0:1], 16, v[0:1]
; VI-NEXT:    s_and_b32 s4, s66, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s66, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s64, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s64, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s62, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[30:31], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s62, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s60, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[28:29], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s60, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s58, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[26:27], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s58, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s56, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[24:25], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s56, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s54, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[22:23], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s54, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s52, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[20:21], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s52, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s37, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s37, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_and_b32 s4, s39, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v5, v7, vcc
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    s_lshl_b32 s4, s39, 16
; VI-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_mov_b32_e32 v1, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v7, v9, vcc
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    s_lshl_b32 s4, s41, 16
; VI-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v3
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_mov_b32_e32 v3, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v9, v11, vcc
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    s_lshl_b32 s4, s43, 16
; VI-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_mov_b32_e32 v5, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v11, v13, vcc
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    s_lshl_b32 s4, s45, 16
; VI-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[16:17], 16, v[15:16]
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_mov_b32_e32 v7, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v13, v15, vcc
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    s_lshl_b32 s4, s47, 16
; VI-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_and_b32 s4, s49, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_mov_b32_e32 v9, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v15, v17, vcc
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    s_lshl_b32 s4, s49, 16
; VI-NEXT:    v_cndmask_b32_e32 v13, v15, v17, vcc
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_and_b32 s4, s51, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_mov_b32_e32 v11, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v17, v19, vcc
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    s_lshl_b32 s4, s51, 16
; VI-NEXT:    v_cndmask_b32_e32 v15, v17, v19, vcc
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v13
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    s_and_b32 s4, s53, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_mov_b32_e32 v13, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v19, v21, vcc
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    s_lshl_b32 s4, s53, 16
; VI-NEXT:    v_cndmask_b32_e32 v17, v19, v21, vcc
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    s_and_b32 s4, s55, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_mov_b32_e32 v15, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v21, v23, vcc
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    s_lshl_b32 s4, s55, 16
; VI-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    s_and_b32 s4, s57, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_mov_b32_e32 v17, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v23, v25, vcc
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    s_lshl_b32 s4, s57, 16
; VI-NEXT:    v_cndmask_b32_e32 v21, v23, v25, vcc
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    s_and_b32 s4, s59, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_mov_b32_e32 v19, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v25, v27, vcc
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    s_lshl_b32 s4, s59, 16
; VI-NEXT:    v_cndmask_b32_e32 v23, v25, v27, vcc
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v21
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    s_and_b32 s4, s61, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_mov_b32_e32 v21, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v27, v29, vcc
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    s_lshl_b32 s4, s61, 16
; VI-NEXT:    v_cndmask_b32_e32 v25, v27, v29, vcc
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    s_and_b32 s4, s63, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_mov_b32_e32 v23, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v29, v31, vcc
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    s_lshl_b32 s4, s63, 16
; VI-NEXT:    v_cndmask_b32_e32 v27, v29, v31, vcc
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    s_and_b32 s4, s65, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v25, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v31, v33, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    v_mov_b32_e32 v27, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    s_lshl_b32 s4, s65, 16
; VI-NEXT:    v_cndmask_b32_e32 v29, v31, v33, vcc
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    s_and_b32 s4, s67, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_mov_b32_e32 v29, v33
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    s_lshl_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v32, v31, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_add_u32_e32 v32, vcc, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x7fff, v32
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v32, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; VI-NEXT:    v_lshrrev_b64 v[31:32], 16, v[31:32]
; VI-NEXT:    s_branch .LBB43_5
; VI-NEXT:  .LBB43_3:
; VI-NEXT:    s_branch .LBB43_2
; VI-NEXT:  .LBB43_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB43_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v35, 15
; VI-NEXT:    v_readlane_b32 s66, v35, 14
; VI-NEXT:    v_readlane_b32 s65, v35, 13
; VI-NEXT:    v_readlane_b32 s64, v35, 12
; VI-NEXT:    v_readlane_b32 s55, v35, 11
; VI-NEXT:    v_readlane_b32 s54, v35, 10
; VI-NEXT:    v_readlane_b32 s53, v35, 9
; VI-NEXT:    v_readlane_b32 s52, v35, 8
; VI-NEXT:    v_readlane_b32 s51, v35, 7
; VI-NEXT:    v_readlane_b32 s50, v35, 6
; VI-NEXT:    v_readlane_b32 s49, v35, 5
; VI-NEXT:    v_readlane_b32 s48, v35, 4
; VI-NEXT:    v_readlane_b32 s39, v35, 3
; VI-NEXT:    v_readlane_b32 s38, v35, 2
; VI-NEXT:    v_readlane_b32 s37, v35, 1
; VI-NEXT:    v_readlane_b32 s36, v35, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v32f32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v36, s36, 0
; GFX9-NEXT:    v_writelane_b32 v36, s37, 1
; GFX9-NEXT:    v_writelane_b32 v36, s38, 2
; GFX9-NEXT:    v_writelane_b32 v36, s39, 3
; GFX9-NEXT:    v_writelane_b32 v36, s48, 4
; GFX9-NEXT:    v_writelane_b32 v36, s49, 5
; GFX9-NEXT:    v_writelane_b32 v36, s50, 6
; GFX9-NEXT:    v_writelane_b32 v36, s51, 7
; GFX9-NEXT:    v_writelane_b32 v36, s52, 8
; GFX9-NEXT:    v_writelane_b32 v36, s53, 9
; GFX9-NEXT:    v_writelane_b32 v36, s54, 10
; GFX9-NEXT:    v_writelane_b32 v36, s55, 11
; GFX9-NEXT:    v_writelane_b32 v36, s64, 12
; GFX9-NEXT:    v_writelane_b32 v36, s65, 13
; GFX9-NEXT:    v_writelane_b32 v36, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v36, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB43_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB43_4
; GFX9-NEXT:  .LBB43_2: ; %cmp.true
; GFX9-NEXT:    s_and_b32 s4, s51, 0xffff0000
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x40c00000
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s51, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    s_and_b32 s4, s50, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_add_f32_e32 v2, s4, v16
; GFX9-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v3, v3, v2
; GFX9-NEXT:    v_add_u32_e32 v3, 0x7fff, v3
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    s_lshl_b32 s4, s50, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT:    v_add_f32_e32 v3, s4, v16
; GFX9-NEXT:    v_mov_b32_e32 v32, 0xffff
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    s_and_b32 s4, s49, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_lshl_or_b32 v15, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_and_b32_sdwa v3, v32, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_lshl_or_b32 v14, v2, 16, v3
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s49, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s48, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s48, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s47, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s47, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s46, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s46, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s45, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s45, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s44, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s44, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s43, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s43, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s42, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s42, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s41, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s41, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s40, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s40, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s39, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s39, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s38, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s38, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v17, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s37, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s37, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v17, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v1
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v17, v18, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s36, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v0
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s36, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v17, v18, vcc
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v17, v32, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s67, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s67, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s66, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v31, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s66, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s65, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v30, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s65, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s64, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v29, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s64, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s63, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v28, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s63, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s62, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v27, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s62, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s61, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v26, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s61, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s60, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s60, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s59, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v24, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s59, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s58, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v23, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s58, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s57, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v22, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s57, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s56, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v21, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s56, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s55, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v20, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s55, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v33, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s54, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v19, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s54, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v33, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v18
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s53, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v17
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s53, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v33, v32, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s52, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v33
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    s_lshl_b32 s4, s52, 16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v16, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v16, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v34, v35, vcc
; GFX9-NEXT:    v_and_b32_sdwa v16, v32, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v16, v32, 16, v16
; GFX9-NEXT:    s_branch .LBB43_5
; GFX9-NEXT:  .LBB43_3:
; GFX9-NEXT:    s_branch .LBB43_2
; GFX9-NEXT:  .LBB43_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB43_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v36, 15
; GFX9-NEXT:    v_readlane_b32 s66, v36, 14
; GFX9-NEXT:    v_readlane_b32 s65, v36, 13
; GFX9-NEXT:    v_readlane_b32 s64, v36, 12
; GFX9-NEXT:    v_readlane_b32 s55, v36, 11
; GFX9-NEXT:    v_readlane_b32 s54, v36, 10
; GFX9-NEXT:    v_readlane_b32 s53, v36, 9
; GFX9-NEXT:    v_readlane_b32 s52, v36, 8
; GFX9-NEXT:    v_readlane_b32 s51, v36, 7
; GFX9-NEXT:    v_readlane_b32 s50, v36, 6
; GFX9-NEXT:    v_readlane_b32 s49, v36, 5
; GFX9-NEXT:    v_readlane_b32 s48, v36, 4
; GFX9-NEXT:    v_readlane_b32 s39, v36, 3
; GFX9-NEXT:    v_readlane_b32 s38, v36, 2
; GFX9-NEXT:    v_readlane_b32 s37, v36, 1
; GFX9-NEXT:    v_readlane_b32 s36, v36, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v32f32_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s36, s0
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-TRUE16-NEXT:    s_mov_b32 s48, s24
; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
; GFX11-TRUE16-NEXT:    s_mov_b32 s39, s3
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-TRUE16-NEXT:    s_mov_b32 s49, s25
; GFX11-TRUE16-NEXT:    s_mov_b32 s38, s2
; GFX11-TRUE16-NEXT:    s_mov_b32 s37, s1
; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-TRUE16-NEXT:    s_mov_b32 s50, s26
; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-TRUE16-NEXT:    s_mov_b32 s51, s27
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-TRUE16-NEXT:    s_mov_b32 s52, s28
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-TRUE16-NEXT:    s_mov_b32 s53, s29
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB43_3
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB43_4
; GFX11-TRUE16-NEXT:  .LBB43_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s51, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s50, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s57, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v0.l
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v2.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v4, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v7, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v16, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v18, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v17 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v16, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v16 :: v_dual_add_nc_u32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v17, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v21, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s37, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v0.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v20 :: v_dual_add_nc_u32 v0, v16, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s36, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v22, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v16.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s67, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v24, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s66, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s65, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v20, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_nc_u32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s64, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s63, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s62, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s61, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v16.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s60, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s58, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v33, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s57, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v33 :: v_dual_add_nc_u32 v19, v19, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v32
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v33, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v20, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s55, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v20, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v35, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s52, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v19, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v33, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v19, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, v35, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v34, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v17.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v36, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v16, v34
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v36, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v35 :: v_dual_add_nc_u32 v35, v37, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v33
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v34, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v37, v38, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v17.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v35, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v32
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v16.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v33.l
; GFX11-TRUE16-NEXT:    s_branch .LBB43_5
; GFX11-TRUE16-NEXT:  .LBB43_3:
; GFX11-TRUE16-NEXT:    s_branch .LBB43_2
; GFX11-TRUE16-NEXT:  .LBB43_4:
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-TRUE16-NEXT:  .LBB43_5: ; %end
; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v32f32_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s36, s0
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-FAKE16-NEXT:    s_mov_b32 s47, s23
; GFX11-FAKE16-NEXT:    s_mov_b32 s46, s22
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-FAKE16-NEXT:    s_mov_b32 s45, s21
; GFX11-FAKE16-NEXT:    s_mov_b32 s44, s20
; GFX11-FAKE16-NEXT:    s_mov_b32 s43, s19
; GFX11-FAKE16-NEXT:    s_mov_b32 s42, s18
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-FAKE16-NEXT:    s_mov_b32 s48, s24
; GFX11-FAKE16-NEXT:    s_mov_b32 s41, s17
; GFX11-FAKE16-NEXT:    s_mov_b32 s40, s16
; GFX11-FAKE16-NEXT:    s_mov_b32 s39, s3
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-FAKE16-NEXT:    s_mov_b32 s49, s25
; GFX11-FAKE16-NEXT:    s_mov_b32 s38, s2
; GFX11-FAKE16-NEXT:    s_mov_b32 s37, s1
; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-FAKE16-NEXT:    s_mov_b32 s50, s26
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-FAKE16-NEXT:    s_mov_b32 s51, s27
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-FAKE16-NEXT:    s_mov_b32 s52, s28
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-FAKE16-NEXT:    s_mov_b32 s53, s29
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB43_3
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB43_4
; GFX11-FAKE16-NEXT:  .LBB43_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s51, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s50, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s49, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s49, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s48, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v7 :: v_dual_add_nc_u32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s47, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s46, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s45, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s44, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s43, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s42, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s41, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s40, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v17 :: v_dual_add_nc_u32 v5, v7, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s39, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s38, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s37, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v4, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s67, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v2, v22 :: v_dual_add_nc_u32 v17, v17, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v22, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s66, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s65, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v22, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s64, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s63, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s62, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s61, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s60, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s58, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v32, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v22, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v19, v21 :: v_dual_and_b32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s57, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v18, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s57, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v22, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v23, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v33, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v35, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v21, v33, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v33, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v22, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s55, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v34, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v33, v34 :: v_dual_add_nc_u32 v33, v35, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v37, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v33, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s52, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v38, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v35, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, v48, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v39
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v48
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v39, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v36, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v37, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v17, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v34, 16, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v32, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v36, 16, v33
; GFX11-FAKE16-NEXT:    s_branch .LBB43_5
; GFX11-FAKE16-NEXT:  .LBB43_3:
; GFX11-FAKE16-NEXT:    s_branch .LBB43_2
; GFX11-FAKE16-NEXT:  .LBB43_4:
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-FAKE16-NEXT:  .LBB43_5: ; %end
; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define <64 x half> @bitcast_v32f32_to_v64f16(<32 x float> %a, i32 %b) {
; SI-LABEL: bitcast_v32f32_to_v64f16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB44_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB44_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB44_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB44_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
; SI-NEXT:    v_or_b32_e32 v0, v0, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v44
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v42
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v61
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v47
; SI-NEXT:    v_or_b32_e32 v3, v3, v44
; SI-NEXT:    v_or_b32_e32 v5, v5, v42
; SI-NEXT:    v_or_b32_e32 v6, v6, v55
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v60
; SI-NEXT:    v_or_b32_e32 v8, v8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v57
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v45
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v41
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v54
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v52
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v53
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v64f16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB44_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; VI-NEXT:  .LBB44_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v64f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB44_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15
; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14
; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13
; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12
; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11
; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10
; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9
; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8
; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7
; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6
; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5
; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4
; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3
; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2
; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30
; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29
; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28
; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27
; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26
; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25
; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24
; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23
; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22
; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21
; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20
; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19
; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18
; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17
; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16
; GFX9-NEXT:  .LBB44_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v64f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB44_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
; GFX11-NEXT:  .LBB44_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define inreg <64 x half> @bitcast_v32f32_to_v64f16_scalar(<32 x float> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32f32_to_v64f16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s44, v18
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_readfirstlane_b32 s7, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    v_readfirstlane_b32 s9, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_readfirstlane_b32 s10, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s15, v7
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s40, v4
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    v_readfirstlane_b32 s42, v2
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB45_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s69, s5, 16
; SI-NEXT:    s_lshr_b32 s68, s7, 16
; SI-NEXT:    s_lshr_b32 s67, s9, 16
; SI-NEXT:    s_lshr_b32 s66, s11, 16
; SI-NEXT:    s_lshr_b32 s65, s13, 16
; SI-NEXT:    s_lshr_b32 s64, s15, 16
; SI-NEXT:    s_lshr_b32 s55, s41, 16
; SI-NEXT:    s_lshr_b32 s54, s43, 16
; SI-NEXT:    s_lshr_b32 s53, s45, 16
; SI-NEXT:    s_lshr_b32 s52, s29, 16
; SI-NEXT:    s_lshr_b32 s51, s27, 16
; SI-NEXT:    s_lshr_b32 s50, s25, 16
; SI-NEXT:    s_lshr_b32 s49, s23, 16
; SI-NEXT:    s_lshr_b32 s48, s21, 16
; SI-NEXT:    s_lshr_b32 s39, s19, 16
; SI-NEXT:    s_lshr_b32 s38, s17, 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB45_4
; SI-NEXT:  .LBB45_2: ; %cmp.true
; SI-NEXT:    v_add_f32_e64 v31, s5, 1.0
; SI-NEXT:    v_add_f32_e64 v30, s4, 1.0
; SI-NEXT:    v_lshr_b64 v[32:33], v[30:31], 16
; SI-NEXT:    v_add_f32_e64 v29, s7, 1.0
; SI-NEXT:    v_add_f32_e64 v28, s6, 1.0
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f32_e64 v17, s43, 1.0
; SI-NEXT:    v_add_f32_e64 v16, s42, 1.0
; SI-NEXT:    v_add_f32_e64 v27, s9, 1.0
; SI-NEXT:    v_add_f32_e64 v26, s8, 1.0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[33:34], v[28:29], 16
; SI-NEXT:    v_add_f32_e64 v15, s45, 1.0
; SI-NEXT:    v_add_f32_e64 v14, s44, 1.0
; SI-NEXT:    v_add_f32_e64 v25, s11, 1.0
; SI-NEXT:    v_add_f32_e64 v24, s10, 1.0
; SI-NEXT:    v_lshr_b64 v[34:35], v[26:27], 16
; SI-NEXT:    v_lshr_b64 v[48:49], v[16:17], 16
; SI-NEXT:    v_add_f32_e64 v13, s29, 1.0
; SI-NEXT:    v_add_f32_e64 v12, s28, 1.0
; SI-NEXT:    v_add_f32_e64 v23, s13, 1.0
; SI-NEXT:    v_add_f32_e64 v22, s12, 1.0
; SI-NEXT:    v_lshr_b64 v[35:36], v[24:25], 16
; SI-NEXT:    v_lshr_b64 v[49:50], v[14:15], 16
; SI-NEXT:    v_add_f32_e64 v11, s27, 1.0
; SI-NEXT:    v_add_f32_e64 v10, s26, 1.0
; SI-NEXT:    v_add_f32_e64 v21, s15, 1.0
; SI-NEXT:    v_add_f32_e64 v20, s14, 1.0
; SI-NEXT:    v_lshr_b64 v[36:37], v[22:23], 16
; SI-NEXT:    v_lshr_b64 v[50:51], v[12:13], 16
; SI-NEXT:    v_add_f32_e64 v9, s25, 1.0
; SI-NEXT:    v_add_f32_e64 v8, s24, 1.0
; SI-NEXT:    v_add_f32_e64 v19, s41, 1.0
; SI-NEXT:    v_add_f32_e64 v18, s40, 1.0
; SI-NEXT:    v_lshr_b64 v[37:38], v[20:21], 16
; SI-NEXT:    v_lshr_b64 v[51:52], v[10:11], 16
; SI-NEXT:    v_add_f32_e64 v3, s19, 1.0
; SI-NEXT:    v_add_f32_e64 v2, s18, 1.0
; SI-NEXT:    v_add_f32_e64 v7, s23, 1.0
; SI-NEXT:    v_add_f32_e64 v6, s22, 1.0
; SI-NEXT:    v_lshr_b64 v[38:39], v[18:19], 16
; SI-NEXT:    v_lshr_b64 v[52:53], v[8:9], 16
; SI-NEXT:    v_add_f32_e64 v1, s17, 1.0
; SI-NEXT:    v_add_f32_e64 v0, s16, 1.0
; SI-NEXT:    v_add_f32_e64 v5, s21, 1.0
; SI-NEXT:    v_add_f32_e64 v4, s20, 1.0
; SI-NEXT:    v_lshr_b64 v[53:54], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[39:40], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[54:55], v[4:5], 16
; SI-NEXT:    v_lshr_b64 v[40:41], v[0:1], 16
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; SI-NEXT:    s_branch .LBB45_5
; SI-NEXT:  .LBB45_3:
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    s_branch .LBB45_2
; SI-NEXT:  .LBB45_4:
; SI-NEXT:    v_mov_b32_e32 v33, s30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_mov_b32_e32 v32, s38
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v62, s39
; SI-NEXT:    v_mov_b32_e32 v61, s48
; SI-NEXT:    v_mov_b32_e32 v60, s49
; SI-NEXT:    v_mov_b32_e32 v59, s50
; SI-NEXT:    v_mov_b32_e32 v58, s51
; SI-NEXT:    v_mov_b32_e32 v57, s52
; SI-NEXT:    v_mov_b32_e32 v56, s53
; SI-NEXT:    v_mov_b32_e32 v47, s54
; SI-NEXT:    v_mov_b32_e32 v46, s55
; SI-NEXT:    v_mov_b32_e32 v45, s64
; SI-NEXT:    v_mov_b32_e32 v44, s65
; SI-NEXT:    v_mov_b32_e32 v43, s66
; SI-NEXT:    v_mov_b32_e32 v42, s67
; SI-NEXT:    v_mov_b32_e32 v41, s68
; SI-NEXT:    v_mov_b32_e32 v55, s69
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v33, s34
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v34, s36
; SI-NEXT:    v_mov_b32_e32 v35, s46
; SI-NEXT:    v_mov_b32_e32 v36, s56
; SI-NEXT:    v_mov_b32_e32 v37, s58
; SI-NEXT:    v_mov_b32_e32 v38, s60
; SI-NEXT:    v_mov_b32_e32 v48, s62
; SI-NEXT:    v_mov_b32_e32 v49, s72
; SI-NEXT:    v_mov_b32_e32 v50, s74
; SI-NEXT:    v_mov_b32_e32 v51, s76
; SI-NEXT:    v_mov_b32_e32 v52, s78
; SI-NEXT:    v_mov_b32_e32 v53, s88
; SI-NEXT:    v_mov_b32_e32 v54, s90
; SI-NEXT:    v_mov_b32_e32 v39, s92
; SI-NEXT:    v_mov_b32_e32 v40, s94
; SI-NEXT:  .LBB45_5: ; %end
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v1, v1, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v39
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v32
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v62
; SI-NEXT:    v_or_b32_e32 v3, v3, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v54
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v32
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v61
; SI-NEXT:    v_or_b32_e32 v5, v5, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v53
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v60
; SI-NEXT:    v_or_b32_e32 v7, v7, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v52
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v32
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v59
; SI-NEXT:    v_or_b32_e32 v9, v9, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v32
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v58
; SI-NEXT:    v_or_b32_e32 v11, v11, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v32
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v57
; SI-NEXT:    v_or_b32_e32 v13, v13, v32
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v49
; SI-NEXT:    v_or_b32_e32 v14, v14, v32
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v56
; SI-NEXT:    v_or_b32_e32 v15, v15, v32
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v48
; SI-NEXT:    v_or_b32_e32 v16, v16, v32
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v47
; SI-NEXT:    v_or_b32_e32 v17, v17, v32
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v38
; SI-NEXT:    v_or_b32_e32 v18, v18, v32
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v46
; SI-NEXT:    v_or_b32_e32 v19, v19, v32
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v37
; SI-NEXT:    v_or_b32_e32 v20, v20, v32
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v45
; SI-NEXT:    v_or_b32_e32 v21, v21, v32
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v36
; SI-NEXT:    v_or_b32_e32 v22, v22, v32
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v44
; SI-NEXT:    v_or_b32_e32 v23, v23, v32
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v35
; SI-NEXT:    v_or_b32_e32 v24, v24, v32
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v43
; SI-NEXT:    v_or_b32_e32 v25, v25, v32
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v34
; SI-NEXT:    v_or_b32_e32 v26, v26, v32
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v27, v27, v32
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; SI-NEXT:    v_or_b32_e32 v0, v0, v40
; SI-NEXT:    v_or_b32_e32 v28, v28, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v41
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v55
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v64f16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s36, 2
; VI-NEXT:    v_writelane_b32 v32, s37, 3
; VI-NEXT:    v_writelane_b32 v32, s38, 4
; VI-NEXT:    v_writelane_b32 v32, s39, 5
; VI-NEXT:    v_writelane_b32 v32, s48, 6
; VI-NEXT:    v_writelane_b32 v32, s49, 7
; VI-NEXT:    v_writelane_b32 v32, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB45_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB45_4
; VI-NEXT:  .LBB45_2: ; %cmp.true
; VI-NEXT:    v_add_f32_e64 v15, s31, 1.0
; VI-NEXT:    v_add_f32_e64 v14, s30, 1.0
; VI-NEXT:    v_add_f32_e64 v13, s29, 1.0
; VI-NEXT:    v_add_f32_e64 v12, s28, 1.0
; VI-NEXT:    v_add_f32_e64 v11, s27, 1.0
; VI-NEXT:    v_add_f32_e64 v10, s26, 1.0
; VI-NEXT:    v_add_f32_e64 v9, s25, 1.0
; VI-NEXT:    v_add_f32_e64 v8, s24, 1.0
; VI-NEXT:    v_add_f32_e64 v7, s23, 1.0
; VI-NEXT:    v_add_f32_e64 v6, s22, 1.0
; VI-NEXT:    v_add_f32_e64 v5, s21, 1.0
; VI-NEXT:    v_add_f32_e64 v4, s20, 1.0
; VI-NEXT:    v_add_f32_e64 v3, s19, 1.0
; VI-NEXT:    v_add_f32_e64 v2, s18, 1.0
; VI-NEXT:    v_add_f32_e64 v1, s17, 1.0
; VI-NEXT:    v_add_f32_e64 v0, s16, 1.0
; VI-NEXT:    v_add_f32_e64 v31, s51, 1.0
; VI-NEXT:    v_add_f32_e64 v30, s50, 1.0
; VI-NEXT:    v_add_f32_e64 v29, s49, 1.0
; VI-NEXT:    v_add_f32_e64 v28, s48, 1.0
; VI-NEXT:    v_add_f32_e64 v27, s47, 1.0
; VI-NEXT:    v_add_f32_e64 v26, s46, 1.0
; VI-NEXT:    v_add_f32_e64 v25, s45, 1.0
; VI-NEXT:    v_add_f32_e64 v24, s44, 1.0
; VI-NEXT:    v_add_f32_e64 v23, s43, 1.0
; VI-NEXT:    v_add_f32_e64 v22, s42, 1.0
; VI-NEXT:    v_add_f32_e64 v21, s41, 1.0
; VI-NEXT:    v_add_f32_e64 v20, s40, 1.0
; VI-NEXT:    v_add_f32_e64 v19, s39, 1.0
; VI-NEXT:    v_add_f32_e64 v18, s38, 1.0
; VI-NEXT:    v_add_f32_e64 v17, s37, 1.0
; VI-NEXT:    v_add_f32_e64 v16, s36, 1.0
; VI-NEXT:    s_branch .LBB45_5
; VI-NEXT:  .LBB45_3:
; VI-NEXT:    s_branch .LBB45_2
; VI-NEXT:  .LBB45_4:
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:  .LBB45_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v32, 9
; VI-NEXT:    v_readlane_b32 s50, v32, 8
; VI-NEXT:    v_readlane_b32 s49, v32, 7
; VI-NEXT:    v_readlane_b32 s48, v32, 6
; VI-NEXT:    v_readlane_b32 s39, v32, 5
; VI-NEXT:    v_readlane_b32 s38, v32, 4
; VI-NEXT:    v_readlane_b32 s37, v32, 3
; VI-NEXT:    v_readlane_b32 s36, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v64f16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB45_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB45_4
; GFX9-NEXT:  .LBB45_2: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e64 v15, s31, 1.0
; GFX9-NEXT:    v_add_f32_e64 v14, s30, 1.0
; GFX9-NEXT:    v_add_f32_e64 v13, s29, 1.0
; GFX9-NEXT:    v_add_f32_e64 v12, s28, 1.0
; GFX9-NEXT:    v_add_f32_e64 v11, s27, 1.0
; GFX9-NEXT:    v_add_f32_e64 v10, s26, 1.0
; GFX9-NEXT:    v_add_f32_e64 v9, s25, 1.0
; GFX9-NEXT:    v_add_f32_e64 v8, s24, 1.0
; GFX9-NEXT:    v_add_f32_e64 v7, s23, 1.0
; GFX9-NEXT:    v_add_f32_e64 v6, s22, 1.0
; GFX9-NEXT:    v_add_f32_e64 v5, s21, 1.0
; GFX9-NEXT:    v_add_f32_e64 v4, s20, 1.0
; GFX9-NEXT:    v_add_f32_e64 v3, s19, 1.0
; GFX9-NEXT:    v_add_f32_e64 v2, s18, 1.0
; GFX9-NEXT:    v_add_f32_e64 v1, s17, 1.0
; GFX9-NEXT:    v_add_f32_e64 v0, s16, 1.0
; GFX9-NEXT:    v_add_f32_e64 v31, s51, 1.0
; GFX9-NEXT:    v_add_f32_e64 v30, s50, 1.0
; GFX9-NEXT:    v_add_f32_e64 v29, s49, 1.0
; GFX9-NEXT:    v_add_f32_e64 v28, s48, 1.0
; GFX9-NEXT:    v_add_f32_e64 v27, s47, 1.0
; GFX9-NEXT:    v_add_f32_e64 v26, s46, 1.0
; GFX9-NEXT:    v_add_f32_e64 v25, s45, 1.0
; GFX9-NEXT:    v_add_f32_e64 v24, s44, 1.0
; GFX9-NEXT:    v_add_f32_e64 v23, s43, 1.0
; GFX9-NEXT:    v_add_f32_e64 v22, s42, 1.0
; GFX9-NEXT:    v_add_f32_e64 v21, s41, 1.0
; GFX9-NEXT:    v_add_f32_e64 v20, s40, 1.0
; GFX9-NEXT:    v_add_f32_e64 v19, s39, 1.0
; GFX9-NEXT:    v_add_f32_e64 v18, s38, 1.0
; GFX9-NEXT:    v_add_f32_e64 v17, s37, 1.0
; GFX9-NEXT:    v_add_f32_e64 v16, s36, 1.0
; GFX9-NEXT:    s_branch .LBB45_5
; GFX9-NEXT:  .LBB45_3:
; GFX9-NEXT:    s_branch .LBB45_2
; GFX9-NEXT:  .LBB45_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:  .LBB45_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v64f16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB45_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB45_4
; GFX11-NEXT:  .LBB45_2: ; %cmp.true
; GFX11-NEXT:    v_add_f32_e64 v15, s27, 1.0
; GFX11-NEXT:    v_add_f32_e64 v14, s26, 1.0
; GFX11-NEXT:    v_add_f32_e64 v13, s25, 1.0
; GFX11-NEXT:    v_add_f32_e64 v12, s24, 1.0
; GFX11-NEXT:    v_add_f32_e64 v11, s23, 1.0
; GFX11-NEXT:    v_add_f32_e64 v10, s22, 1.0
; GFX11-NEXT:    v_add_f32_e64 v9, s21, 1.0
; GFX11-NEXT:    v_add_f32_e64 v8, s20, 1.0
; GFX11-NEXT:    v_add_f32_e64 v7, s19, 1.0
; GFX11-NEXT:    v_add_f32_e64 v6, s18, 1.0
; GFX11-NEXT:    v_add_f32_e64 v5, s17, 1.0
; GFX11-NEXT:    v_add_f32_e64 v4, s16, 1.0
; GFX11-NEXT:    v_add_f32_e64 v3, s15, 1.0
; GFX11-NEXT:    v_add_f32_e64 v2, s14, 1.0
; GFX11-NEXT:    v_add_f32_e64 v1, s13, 1.0
; GFX11-NEXT:    v_add_f32_e64 v0, s12, 1.0
; GFX11-NEXT:    v_add_f32_e64 v31, s51, 1.0
; GFX11-NEXT:    v_add_f32_e64 v30, s50, 1.0
; GFX11-NEXT:    v_add_f32_e64 v29, s49, 1.0
; GFX11-NEXT:    v_add_f32_e64 v28, s48, 1.0
; GFX11-NEXT:    v_add_f32_e64 v27, s47, 1.0
; GFX11-NEXT:    v_add_f32_e64 v26, s46, 1.0
; GFX11-NEXT:    v_add_f32_e64 v25, s45, 1.0
; GFX11-NEXT:    v_add_f32_e64 v24, s44, 1.0
; GFX11-NEXT:    v_add_f32_e64 v23, s43, 1.0
; GFX11-NEXT:    v_add_f32_e64 v22, s42, 1.0
; GFX11-NEXT:    v_add_f32_e64 v21, s41, 1.0
; GFX11-NEXT:    v_add_f32_e64 v20, s40, 1.0
; GFX11-NEXT:    v_add_f32_e64 v19, s39, 1.0
; GFX11-NEXT:    v_add_f32_e64 v18, s38, 1.0
; GFX11-NEXT:    v_add_f32_e64 v17, s37, 1.0
; GFX11-NEXT:    v_add_f32_e64 v16, s36, 1.0
; GFX11-NEXT:    s_branch .LBB45_5
; GFX11-NEXT:  .LBB45_3:
; GFX11-NEXT:    s_branch .LBB45_2
; GFX11-NEXT:  .LBB45_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:  .LBB45_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define <32 x float> @bitcast_v64f16_to_v32f32(<64 x half> %a, i32 %b) {
; SI-LABEL: bitcast_v64f16_to_v32f32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v30
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v29
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v27
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v25
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v24
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v23
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v22
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v21
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v20
; SI-NEXT:    v_mov_b32_e32 v52, v18
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
; SI-NEXT:    v_mov_b32_e32 v53, v17
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v52
; SI-NEXT:    v_mov_b32_e32 v54, v16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v53
; SI-NEXT:    v_mov_b32_e32 v55, v15
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v54
; SI-NEXT:    v_mov_b32_e32 v40, v14
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v55
; SI-NEXT:    v_mov_b32_e32 v41, v13
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v40
; SI-NEXT:    v_mov_b32_e32 v42, v12
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v28
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v42
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v61
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v62
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB46_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v63
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v61
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v32
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v36
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v49
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v33
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v48
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v35
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v39
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v38
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v37
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v42
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v41
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v40
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v55
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v53
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v52
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB46_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB46_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, v63
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v50
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v62
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v61
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_or_b32_e32 v1, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v59
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v58
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v57
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v56
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v49
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v47
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v46
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v33
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v45
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v44
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v48
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v43
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v42
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v35
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v41
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v40
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v55
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v54
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v34
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v53
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v19, v52
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v38
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v37
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v51
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:  .LBB46_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v32f32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB46_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x200
; VI-NEXT:    v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v15, 0x200, v15
; VI-NEXT:    v_or_b32_e32 v15, v15, v33
; VI-NEXT:    v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v14, 0x200, v14
; VI-NEXT:    v_or_b32_e32 v14, v14, v33
; VI-NEXT:    v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v13, 0x200, v13
; VI-NEXT:    v_or_b32_e32 v13, v13, v33
; VI-NEXT:    v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v12, 0x200, v12
; VI-NEXT:    v_or_b32_e32 v12, v12, v33
; VI-NEXT:    v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v11, 0x200, v11
; VI-NEXT:    v_or_b32_e32 v11, v11, v33
; VI-NEXT:    v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v10, 0x200, v10
; VI-NEXT:    v_or_b32_e32 v10, v10, v33
; VI-NEXT:    v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v9, 0x200, v9
; VI-NEXT:    v_or_b32_e32 v9, v9, v33
; VI-NEXT:    v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v8, 0x200, v8
; VI-NEXT:    v_or_b32_e32 v8, v8, v33
; VI-NEXT:    v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v7, 0x200, v7
; VI-NEXT:    v_or_b32_e32 v7, v7, v33
; VI-NEXT:    v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v6, 0x200, v6
; VI-NEXT:    v_or_b32_e32 v6, v6, v33
; VI-NEXT:    v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v5, 0x200, v5
; VI-NEXT:    v_or_b32_e32 v5, v5, v33
; VI-NEXT:    v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v4, 0x200, v4
; VI-NEXT:    v_or_b32_e32 v4, v4, v33
; VI-NEXT:    v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, 0x200, v3
; VI-NEXT:    v_or_b32_e32 v3, v3, v33
; VI-NEXT:    v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v2, 0x200, v2
; VI-NEXT:    v_or_b32_e32 v2, v2, v33
; VI-NEXT:    v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, 0x200, v1
; VI-NEXT:    v_or_b32_e32 v1, v1, v33
; VI-NEXT:    v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, 0x200, v0
; VI-NEXT:    v_or_b32_e32 v0, v0, v33
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v31, 0x200, v31
; VI-NEXT:    v_or_b32_e32 v31, v31, v33
; VI-NEXT:    v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v30, 0x200, v30
; VI-NEXT:    v_or_b32_e32 v30, v30, v33
; VI-NEXT:    v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v29, 0x200, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v28, 0x200, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v33
; VI-NEXT:    v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v27, 0x200, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v33
; VI-NEXT:    v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v26, 0x200, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v33
; VI-NEXT:    v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v25, 0x200, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v33
; VI-NEXT:    v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v24, 0x200, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v33
; VI-NEXT:    v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v23, 0x200, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v33
; VI-NEXT:    v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v22, 0x200, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v33
; VI-NEXT:    v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v21, 0x200, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v33
; VI-NEXT:    v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v20, 0x200, v20
; VI-NEXT:    v_or_b32_e32 v20, v20, v33
; VI-NEXT:    v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v19, 0x200, v19
; VI-NEXT:    v_or_b32_e32 v19, v19, v33
; VI-NEXT:    v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, 0x200, v18
; VI-NEXT:    v_or_b32_e32 v18, v18, v33
; VI-NEXT:    v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, 0x200, v17
; VI-NEXT:    v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, 0x200, v16
; VI-NEXT:    v_or_b32_e32 v17, v17, v33
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB46_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v32f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB46_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_movk_i32 s6, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB46_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v32f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB46_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-NEXT:  .LBB46_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define inreg <32 x float> @bitcast_v64f16_to_v32f32_scalar(<64 x half> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64f16_to_v32f32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v34, s30, 0
; SI-NEXT:    v_writelane_b32 v34, s31, 1
; SI-NEXT:    v_writelane_b32 v34, s34, 2
; SI-NEXT:    v_writelane_b32 v34, s35, 3
; SI-NEXT:    v_writelane_b32 v34, s36, 4
; SI-NEXT:    v_writelane_b32 v34, s37, 5
; SI-NEXT:    v_writelane_b32 v34, s38, 6
; SI-NEXT:    v_writelane_b32 v34, s39, 7
; SI-NEXT:    v_writelane_b32 v34, s48, 8
; SI-NEXT:    v_writelane_b32 v34, s49, 9
; SI-NEXT:    v_writelane_b32 v34, s50, 10
; SI-NEXT:    v_writelane_b32 v34, s51, 11
; SI-NEXT:    v_writelane_b32 v34, s52, 12
; SI-NEXT:    v_writelane_b32 v34, s53, 13
; SI-NEXT:    v_writelane_b32 v34, s54, 14
; SI-NEXT:    v_writelane_b32 v34, s55, 15
; SI-NEXT:    v_writelane_b32 v34, s64, 16
; SI-NEXT:    v_writelane_b32 v34, s65, 17
; SI-NEXT:    v_writelane_b32 v34, s66, 18
; SI-NEXT:    v_writelane_b32 v34, s67, 19
; SI-NEXT:    v_writelane_b32 v34, s68, 20
; SI-NEXT:    v_writelane_b32 v34, s69, 21
; SI-NEXT:    v_writelane_b32 v34, s70, 22
; SI-NEXT:    v_writelane_b32 v34, s71, 23
; SI-NEXT:    v_writelane_b32 v34, s80, 24
; SI-NEXT:    v_writelane_b32 v34, s81, 25
; SI-NEXT:    v_writelane_b32 v34, s82, 26
; SI-NEXT:    v_writelane_b32 v34, s83, 27
; SI-NEXT:    v_writelane_b32 v34, s84, 28
; SI-NEXT:    v_writelane_b32 v34, s85, 29
; SI-NEXT:    v_writelane_b32 v34, s86, 30
; SI-NEXT:    v_writelane_b32 v34, s87, 31
; SI-NEXT:    v_writelane_b32 v34, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_writelane_b32 v34, s97, 33
; SI-NEXT:    s_lshr_b32 vcc_lo, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s8, v16
; SI-NEXT:    ; implicit-def: $vgpr35 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v34, s98, 34
; SI-NEXT:    s_lshr_b32 vcc_hi, s8, 16
; SI-NEXT:    v_readfirstlane_b32 s10, v15
; SI-NEXT:    v_readfirstlane_b32 s12, v14
; SI-NEXT:    v_readfirstlane_b32 s14, v13
; SI-NEXT:    v_readfirstlane_b32 s72, v12
; SI-NEXT:    v_readfirstlane_b32 s74, v11
; SI-NEXT:    v_readfirstlane_b32 s76, v10
; SI-NEXT:    v_readfirstlane_b32 s78, v9
; SI-NEXT:    v_readfirstlane_b32 s88, v8
; SI-NEXT:    v_readfirstlane_b32 s90, v7
; SI-NEXT:    v_readfirstlane_b32 s92, v6
; SI-NEXT:    v_readfirstlane_b32 s94, v5
; SI-NEXT:    v_readfirstlane_b32 s30, v4
; SI-NEXT:    v_readfirstlane_b32 s34, v3
; SI-NEXT:    v_readfirstlane_b32 s68, v2
; SI-NEXT:    v_readfirstlane_b32 s80, v1
; SI-NEXT:    v_readfirstlane_b32 s83, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v35, vcc_lo, 0
; SI-NEXT:    v_writelane_b32 v34, s99, 35
; SI-NEXT:    s_lshr_b32 s69, s29, 16
; SI-NEXT:    s_lshr_b32 s71, s28, 16
; SI-NEXT:    s_lshr_b32 s82, s27, 16
; SI-NEXT:    s_lshr_b32 s85, s26, 16
; SI-NEXT:    s_lshr_b32 s86, s25, 16
; SI-NEXT:    s_lshr_b32 s87, s24, 16
; SI-NEXT:    s_lshr_b32 s96, s23, 16
; SI-NEXT:    s_lshr_b32 s97, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s7, s19, 16
; SI-NEXT:    s_lshr_b32 s9, s18, 16
; SI-NEXT:    s_lshr_b32 s11, s17, 16
; SI-NEXT:    s_lshr_b32 s13, s16, 16
; SI-NEXT:    s_lshr_b32 s65, s10, 16
; SI-NEXT:    s_lshr_b32 s64, s12, 16
; SI-NEXT:    s_lshr_b32 s15, s14, 16
; SI-NEXT:    s_lshr_b32 s73, s72, 16
; SI-NEXT:    s_lshr_b32 s75, s74, 16
; SI-NEXT:    s_lshr_b32 s77, s76, 16
; SI-NEXT:    s_lshr_b32 s79, s78, 16
; SI-NEXT:    s_lshr_b32 s89, s88, 16
; SI-NEXT:    s_lshr_b32 s91, s90, 16
; SI-NEXT:    s_lshr_b32 s93, s92, 16
; SI-NEXT:    s_lshr_b32 s95, s94, 16
; SI-NEXT:    s_lshr_b32 s31, s30, 16
; SI-NEXT:    s_lshr_b32 s35, s34, 16
; SI-NEXT:    s_lshr_b32 s70, s68, 16
; SI-NEXT:    s_lshr_b32 s81, s80, 16
; SI-NEXT:    s_lshr_b32 s84, s83, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v35, vcc_hi, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v35, s65, 2
; SI-NEXT:    v_writelane_b32 v35, s64, 3
; SI-NEXT:    s_cbranch_scc0 .LBB47_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s13, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s9, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s7, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s11, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s97, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s86, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s34, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s30, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s31, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s92, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s90, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s91, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s88, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s89, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s78, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s79, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s76, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s77, 16
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s74, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s75, 16
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s72, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s73, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s15, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s64, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s65, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB47_4
; SI-NEXT:  .LBB47_2: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s13
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s16
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s11
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s17
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s18
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s9
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v1, v3, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s7
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s19
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s99
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s20
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s98
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s21
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s97
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v7, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s22
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s96
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s23
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s87
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s24
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_or_b32_e32 v7, v9, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s86
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v10, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s25
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s85
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s26
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s82
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s27
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_or_b32_e32 v10, v12, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s71
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v13, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s28
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s69
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s29
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s84
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s83
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_or_b32_e32 v13, v15, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s81
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s80
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s70
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s68
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s35
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s34
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_or_b32_e32 v16, v18, v16
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s31
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s30
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s95
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s94
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s93
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s92
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s91
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s90
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s89
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s88
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s79
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s78
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s77
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s76
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s75
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s74
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s73
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s72
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s15
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_readlane_b32 s4, v35, 3
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s14
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_readlane_b32 s4, v35, 2
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s12
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s10
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_readlane_b32 s4, v35, 1
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s4
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_readlane_b32 s4, v35, 0
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s8
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v33, s6
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:    s_branch .LBB47_5
; SI-NEXT:  .LBB47_3:
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB47_2
; SI-NEXT:  .LBB47_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB47_5: ; %end
; SI-NEXT:    v_readlane_b32 s99, v34, 35
; SI-NEXT:    v_readlane_b32 s98, v34, 34
; SI-NEXT:    v_readlane_b32 s97, v34, 33
; SI-NEXT:    v_readlane_b32 s96, v34, 32
; SI-NEXT:    v_readlane_b32 s87, v34, 31
; SI-NEXT:    v_readlane_b32 s86, v34, 30
; SI-NEXT:    v_readlane_b32 s85, v34, 29
; SI-NEXT:    v_readlane_b32 s84, v34, 28
; SI-NEXT:    v_readlane_b32 s83, v34, 27
; SI-NEXT:    v_readlane_b32 s82, v34, 26
; SI-NEXT:    v_readlane_b32 s81, v34, 25
; SI-NEXT:    v_readlane_b32 s80, v34, 24
; SI-NEXT:    v_readlane_b32 s71, v34, 23
; SI-NEXT:    v_readlane_b32 s70, v34, 22
; SI-NEXT:    v_readlane_b32 s69, v34, 21
; SI-NEXT:    v_readlane_b32 s68, v34, 20
; SI-NEXT:    v_readlane_b32 s67, v34, 19
; SI-NEXT:    v_readlane_b32 s66, v34, 18
; SI-NEXT:    v_readlane_b32 s65, v34, 17
; SI-NEXT:    v_readlane_b32 s64, v34, 16
; SI-NEXT:    v_readlane_b32 s55, v34, 15
; SI-NEXT:    v_readlane_b32 s54, v34, 14
; SI-NEXT:    v_readlane_b32 s53, v34, 13
; SI-NEXT:    v_readlane_b32 s52, v34, 12
; SI-NEXT:    v_readlane_b32 s51, v34, 11
; SI-NEXT:    v_readlane_b32 s50, v34, 10
; SI-NEXT:    v_readlane_b32 s49, v34, 9
; SI-NEXT:    v_readlane_b32 s48, v34, 8
; SI-NEXT:    v_readlane_b32 s39, v34, 7
; SI-NEXT:    v_readlane_b32 s38, v34, 6
; SI-NEXT:    v_readlane_b32 s37, v34, 5
; SI-NEXT:    v_readlane_b32 s36, v34, 4
; SI-NEXT:    v_readlane_b32 s35, v34, 3
; SI-NEXT:    v_readlane_b32 s34, v34, 2
; SI-NEXT:    v_readlane_b32 s31, v34, 1
; SI-NEXT:    v_readlane_b32 s30, v34, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v32f32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v33, s36, 0
; VI-NEXT:    v_writelane_b32 v33, s37, 1
; VI-NEXT:    v_writelane_b32 v33, s38, 2
; VI-NEXT:    v_writelane_b32 v33, s39, 3
; VI-NEXT:    v_writelane_b32 v33, s48, 4
; VI-NEXT:    v_writelane_b32 v33, s49, 5
; VI-NEXT:    v_writelane_b32 v33, s50, 6
; VI-NEXT:    v_writelane_b32 v33, s51, 7
; VI-NEXT:    v_writelane_b32 v33, s52, 8
; VI-NEXT:    v_writelane_b32 v33, s53, 9
; VI-NEXT:    v_writelane_b32 v33, s54, 10
; VI-NEXT:    v_writelane_b32 v33, s55, 11
; VI-NEXT:    v_writelane_b32 v33, s64, 12
; VI-NEXT:    v_writelane_b32 v33, s65, 13
; VI-NEXT:    v_writelane_b32 v33, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v33, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB47_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB47_4
; VI-NEXT:  .LBB47_2: ; %cmp.true
; VI-NEXT:    s_lshr_b32 s4, s51, 16
; VI-NEXT:    v_mov_b32_e32 v16, 0x200
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    s_lshr_b32 s4, s50, 16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s51, v16
; VI-NEXT:    v_mov_b32_e32 v2, s4
; VI-NEXT:    s_lshr_b32 s4, s49, 16
; VI-NEXT:    v_or_b32_e32 v15, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s49, v16
; VI-NEXT:    s_lshr_b32 s4, s48, 16
; VI-NEXT:    v_or_b32_e32 v13, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s48, v16
; VI-NEXT:    s_lshr_b32 s4, s47, 16
; VI-NEXT:    v_or_b32_e32 v12, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s47, v16
; VI-NEXT:    s_lshr_b32 s4, s46, 16
; VI-NEXT:    v_or_b32_e32 v11, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s46, v16
; VI-NEXT:    s_lshr_b32 s4, s45, 16
; VI-NEXT:    v_or_b32_e32 v10, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s45, v16
; VI-NEXT:    s_lshr_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v9, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s44, v16
; VI-NEXT:    s_lshr_b32 s4, s43, 16
; VI-NEXT:    v_or_b32_e32 v8, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s43, v16
; VI-NEXT:    s_lshr_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v7, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s42, v16
; VI-NEXT:    s_lshr_b32 s4, s41, 16
; VI-NEXT:    v_or_b32_e32 v6, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s41, v16
; VI-NEXT:    s_lshr_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v5, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s40, v16
; VI-NEXT:    s_lshr_b32 s4, s39, 16
; VI-NEXT:    v_or_b32_e32 v4, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v2, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, s50, v16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s39, v16
; VI-NEXT:    s_lshr_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v14, v3, v2
; VI-NEXT:    v_or_b32_e32 v3, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s38, v16
; VI-NEXT:    s_lshr_b32 s4, s37, 16
; VI-NEXT:    v_or_b32_e32 v2, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s37, v16
; VI-NEXT:    s_lshr_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v1, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, s36, v16
; VI-NEXT:    s_lshr_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v0, v17, v0
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s67, v16
; VI-NEXT:    s_lshr_b32 s4, s66, 16
; VI-NEXT:    v_or_b32_e32 v31, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s66, v16
; VI-NEXT:    s_lshr_b32 s4, s65, 16
; VI-NEXT:    v_or_b32_e32 v30, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s65, v16
; VI-NEXT:    s_lshr_b32 s4, s64, 16
; VI-NEXT:    v_or_b32_e32 v29, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s64, v16
; VI-NEXT:    s_lshr_b32 s4, s63, 16
; VI-NEXT:    v_or_b32_e32 v28, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s63, v16
; VI-NEXT:    s_lshr_b32 s4, s62, 16
; VI-NEXT:    v_or_b32_e32 v27, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s62, v16
; VI-NEXT:    s_lshr_b32 s4, s61, 16
; VI-NEXT:    v_or_b32_e32 v26, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s61, v16
; VI-NEXT:    s_lshr_b32 s4, s60, 16
; VI-NEXT:    v_or_b32_e32 v25, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s60, v16
; VI-NEXT:    s_lshr_b32 s4, s59, 16
; VI-NEXT:    v_or_b32_e32 v24, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s59, v16
; VI-NEXT:    s_lshr_b32 s4, s58, 16
; VI-NEXT:    v_or_b32_e32 v23, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s58, v16
; VI-NEXT:    s_lshr_b32 s4, s57, 16
; VI-NEXT:    v_or_b32_e32 v22, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s57, v16
; VI-NEXT:    s_lshr_b32 s4, s56, 16
; VI-NEXT:    v_or_b32_e32 v21, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s56, v16
; VI-NEXT:    s_lshr_b32 s4, s55, 16
; VI-NEXT:    v_or_b32_e32 v20, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s55, v16
; VI-NEXT:    s_lshr_b32 s4, s54, 16
; VI-NEXT:    v_or_b32_e32 v19, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s54, v16
; VI-NEXT:    s_lshr_b32 s4, s53, 16
; VI-NEXT:    v_or_b32_e32 v18, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v32, s53, v16
; VI-NEXT:    s_lshr_b32 s4, s52, 16
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_mov_b32_e32 v32, s4
; VI-NEXT:    v_add_f16_sdwa v32, v32, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, s52, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:    s_branch .LBB47_5
; VI-NEXT:  .LBB47_3:
; VI-NEXT:    s_branch .LBB47_2
; VI-NEXT:  .LBB47_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB47_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v33, 15
; VI-NEXT:    v_readlane_b32 s66, v33, 14
; VI-NEXT:    v_readlane_b32 s65, v33, 13
; VI-NEXT:    v_readlane_b32 s64, v33, 12
; VI-NEXT:    v_readlane_b32 s55, v33, 11
; VI-NEXT:    v_readlane_b32 s54, v33, 10
; VI-NEXT:    v_readlane_b32 s53, v33, 9
; VI-NEXT:    v_readlane_b32 s52, v33, 8
; VI-NEXT:    v_readlane_b32 s51, v33, 7
; VI-NEXT:    v_readlane_b32 s50, v33, 6
; VI-NEXT:    v_readlane_b32 s49, v33, 5
; VI-NEXT:    v_readlane_b32 s48, v33, 4
; VI-NEXT:    v_readlane_b32 s39, v33, 3
; VI-NEXT:    v_readlane_b32 s38, v33, 2
; VI-NEXT:    v_readlane_b32 s37, v33, 1
; VI-NEXT:    v_readlane_b32 s36, v33, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v32f32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB47_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB47_4
; GFX9-NEXT:  .LBB47_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, s51, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, s50, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, s49, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, s48, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, s47, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, s46, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, s45, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, s44, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, s43, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, s42, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, s41, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, s40, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, s39, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, s38, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, s37, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, s36, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v31, s67, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, s66, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, s65, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, s64, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, s63, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, s62, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, s61, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, s60, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, s59, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, s58, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, s57, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, s56, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, s55, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, s54, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, s53, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, s52, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB47_5
; GFX9-NEXT:  .LBB47_3:
; GFX9-NEXT:    s_branch .LBB47_2
; GFX9-NEXT:  .LBB47_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB47_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v32f32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB47_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB47_4
; GFX11-NEXT:  .LBB47_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s51 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s50 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s49 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s48 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s47 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s46 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s45 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s44 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s43 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s42 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s41 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s40 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s39 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s38 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s37 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s36 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, s67 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, s66 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s65 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s64 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, s63 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, s62 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, s61 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s60 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s59 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, s58 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s57 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s56 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s55 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s54 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s53 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s52 op_sel_hi:[0,1]
; GFX11-NEXT:    s_branch .LBB47_5
; GFX11-NEXT:  .LBB47_3:
; GFX11-NEXT:    s_branch .LBB47_2
; GFX11-NEXT:  .LBB47_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB47_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define <64 x i16> @bitcast_v32f32_to_v64i16(<32 x float> %a, i32 %b) {
; SI-LABEL: bitcast_v32f32_to_v64i16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB48_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB48_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB48_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v53, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v55, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v42, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB48_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
; SI-NEXT:    v_or_b32_e32 v0, v0, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v44
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v42
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v61
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v47
; SI-NEXT:    v_or_b32_e32 v3, v3, v44
; SI-NEXT:    v_or_b32_e32 v5, v5, v42
; SI-NEXT:    v_or_b32_e32 v6, v6, v55
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v60
; SI-NEXT:    v_or_b32_e32 v8, v8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v57
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v45
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v41
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v54
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v52
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v53
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v64i16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB48_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_f32_e32 v15, 1.0, v15
; VI-NEXT:    v_add_f32_e32 v14, 1.0, v14
; VI-NEXT:    v_add_f32_e32 v13, 1.0, v13
; VI-NEXT:    v_add_f32_e32 v12, 1.0, v12
; VI-NEXT:    v_add_f32_e32 v11, 1.0, v11
; VI-NEXT:    v_add_f32_e32 v10, 1.0, v10
; VI-NEXT:    v_add_f32_e32 v9, 1.0, v9
; VI-NEXT:    v_add_f32_e32 v8, 1.0, v8
; VI-NEXT:    v_add_f32_e32 v7, 1.0, v7
; VI-NEXT:    v_add_f32_e32 v6, 1.0, v6
; VI-NEXT:    v_add_f32_e32 v5, 1.0, v5
; VI-NEXT:    v_add_f32_e32 v4, 1.0, v4
; VI-NEXT:    v_add_f32_e32 v3, 1.0, v3
; VI-NEXT:    v_add_f32_e32 v2, 1.0, v2
; VI-NEXT:    v_add_f32_e32 v1, 1.0, v1
; VI-NEXT:    v_add_f32_e32 v0, 1.0, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f32_e32 v31, 1.0, v31
; VI-NEXT:    v_add_f32_e32 v30, 1.0, v30
; VI-NEXT:    v_add_f32_e32 v29, 1.0, v29
; VI-NEXT:    v_add_f32_e32 v28, 1.0, v28
; VI-NEXT:    v_add_f32_e32 v27, 1.0, v27
; VI-NEXT:    v_add_f32_e32 v26, 1.0, v26
; VI-NEXT:    v_add_f32_e32 v25, 1.0, v25
; VI-NEXT:    v_add_f32_e32 v24, 1.0, v24
; VI-NEXT:    v_add_f32_e32 v23, 1.0, v23
; VI-NEXT:    v_add_f32_e32 v22, 1.0, v22
; VI-NEXT:    v_add_f32_e32 v21, 1.0, v21
; VI-NEXT:    v_add_f32_e32 v20, 1.0, v20
; VI-NEXT:    v_add_f32_e32 v19, 1.0, v19
; VI-NEXT:    v_add_f32_e32 v18, 1.0, v18
; VI-NEXT:    v_add_f32_e32 v17, 1.0, v17
; VI-NEXT:    v_add_f32_e32 v16, 1.0, v16
; VI-NEXT:  .LBB48_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v64i16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB48_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e32 v15, 1.0, v15
; GFX9-NEXT:    v_add_f32_e32 v14, 1.0, v14
; GFX9-NEXT:    v_add_f32_e32 v13, 1.0, v13
; GFX9-NEXT:    v_add_f32_e32 v12, 1.0, v12
; GFX9-NEXT:    v_add_f32_e32 v11, 1.0, v11
; GFX9-NEXT:    v_add_f32_e32 v10, 1.0, v10
; GFX9-NEXT:    v_add_f32_e32 v9, 1.0, v9
; GFX9-NEXT:    v_add_f32_e32 v8, 1.0, v8
; GFX9-NEXT:    v_add_f32_e32 v7, 1.0, v7
; GFX9-NEXT:    v_add_f32_e32 v6, 1.0, v6
; GFX9-NEXT:    v_add_f32_e32 v5, 1.0, v5
; GFX9-NEXT:    v_add_f32_e32 v4, 1.0, v4
; GFX9-NEXT:    v_add_f32_e32 v3, 1.0, v3
; GFX9-NEXT:    v_add_f32_e32 v2, 1.0, v2
; GFX9-NEXT:    v_add_f32_e32 v1, 1.0, v1
; GFX9-NEXT:    v_add_f32_e32 v0, 1.0, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f32_e32 v31, 1.0, v31
; GFX9-NEXT:    v_add_f32_e32 v30, 1.0, v30
; GFX9-NEXT:    v_add_f32_e32 v29, 1.0, v29
; GFX9-NEXT:    v_add_f32_e32 v28, 1.0, v28
; GFX9-NEXT:    v_add_f32_e32 v27, 1.0, v27
; GFX9-NEXT:    v_add_f32_e32 v26, 1.0, v26
; GFX9-NEXT:    v_add_f32_e32 v25, 1.0, v25
; GFX9-NEXT:    v_add_f32_e32 v24, 1.0, v24
; GFX9-NEXT:    v_add_f32_e32 v23, 1.0, v23
; GFX9-NEXT:    v_add_f32_e32 v22, 1.0, v22
; GFX9-NEXT:    v_add_f32_e32 v21, 1.0, v21
; GFX9-NEXT:    v_add_f32_e32 v20, 1.0, v20
; GFX9-NEXT:    v_add_f32_e32 v19, 1.0, v19
; GFX9-NEXT:    v_add_f32_e32 v18, 1.0, v18
; GFX9-NEXT:    v_add_f32_e32 v17, 1.0, v17
; GFX9-NEXT:    v_add_f32_e32 v16, 1.0, v16
; GFX9-NEXT:  .LBB48_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v64i16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB48_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_dual_add_f32 v15, 1.0, v15 :: v_dual_add_f32 v14, 1.0, v14
; GFX11-NEXT:    v_dual_add_f32 v13, 1.0, v13 :: v_dual_add_f32 v12, 1.0, v12
; GFX11-NEXT:    v_dual_add_f32 v11, 1.0, v11 :: v_dual_add_f32 v10, 1.0, v10
; GFX11-NEXT:    v_dual_add_f32 v9, 1.0, v9 :: v_dual_add_f32 v8, 1.0, v8
; GFX11-NEXT:    v_dual_add_f32 v7, 1.0, v7 :: v_dual_add_f32 v6, 1.0, v6
; GFX11-NEXT:    v_dual_add_f32 v5, 1.0, v5 :: v_dual_add_f32 v4, 1.0, v4
; GFX11-NEXT:    v_dual_add_f32 v3, 1.0, v3 :: v_dual_add_f32 v2, 1.0, v2
; GFX11-NEXT:    v_dual_add_f32 v1, 1.0, v1 :: v_dual_add_f32 v0, 1.0, v0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_dual_add_f32 v31, 1.0, v31 :: v_dual_add_f32 v30, 1.0, v30
; GFX11-NEXT:    v_dual_add_f32 v29, 1.0, v29 :: v_dual_add_f32 v28, 1.0, v28
; GFX11-NEXT:    v_dual_add_f32 v27, 1.0, v27 :: v_dual_add_f32 v26, 1.0, v26
; GFX11-NEXT:    v_dual_add_f32 v25, 1.0, v25 :: v_dual_add_f32 v24, 1.0, v24
; GFX11-NEXT:    v_dual_add_f32 v23, 1.0, v23 :: v_dual_add_f32 v22, 1.0, v22
; GFX11-NEXT:    v_dual_add_f32 v21, 1.0, v21 :: v_dual_add_f32 v20, 1.0, v20
; GFX11-NEXT:    v_dual_add_f32 v19, 1.0, v19 :: v_dual_add_f32 v18, 1.0, v18
; GFX11-NEXT:    v_dual_add_f32 v17, 1.0, v17 :: v_dual_add_f32 v16, 1.0, v16
; GFX11-NEXT:  .LBB48_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define inreg <64 x i16> @bitcast_v32f32_to_v64i16_scalar(<32 x float> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v32f32_to_v64i16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s44, v18
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_readfirstlane_b32 s7, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    v_readfirstlane_b32 s9, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_readfirstlane_b32 s10, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s15, v7
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s40, v4
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    v_readfirstlane_b32 s42, v2
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB49_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s69, s5, 16
; SI-NEXT:    s_lshr_b32 s68, s7, 16
; SI-NEXT:    s_lshr_b32 s67, s9, 16
; SI-NEXT:    s_lshr_b32 s66, s11, 16
; SI-NEXT:    s_lshr_b32 s65, s13, 16
; SI-NEXT:    s_lshr_b32 s64, s15, 16
; SI-NEXT:    s_lshr_b32 s55, s41, 16
; SI-NEXT:    s_lshr_b32 s54, s43, 16
; SI-NEXT:    s_lshr_b32 s53, s45, 16
; SI-NEXT:    s_lshr_b32 s52, s29, 16
; SI-NEXT:    s_lshr_b32 s51, s27, 16
; SI-NEXT:    s_lshr_b32 s50, s25, 16
; SI-NEXT:    s_lshr_b32 s49, s23, 16
; SI-NEXT:    s_lshr_b32 s48, s21, 16
; SI-NEXT:    s_lshr_b32 s39, s19, 16
; SI-NEXT:    s_lshr_b32 s38, s17, 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB49_4
; SI-NEXT:  .LBB49_2: ; %cmp.true
; SI-NEXT:    v_add_f32_e64 v31, s5, 1.0
; SI-NEXT:    v_add_f32_e64 v30, s4, 1.0
; SI-NEXT:    v_lshr_b64 v[32:33], v[30:31], 16
; SI-NEXT:    v_add_f32_e64 v29, s7, 1.0
; SI-NEXT:    v_add_f32_e64 v28, s6, 1.0
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f32_e64 v17, s43, 1.0
; SI-NEXT:    v_add_f32_e64 v16, s42, 1.0
; SI-NEXT:    v_add_f32_e64 v27, s9, 1.0
; SI-NEXT:    v_add_f32_e64 v26, s8, 1.0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[33:34], v[28:29], 16
; SI-NEXT:    v_add_f32_e64 v15, s45, 1.0
; SI-NEXT:    v_add_f32_e64 v14, s44, 1.0
; SI-NEXT:    v_add_f32_e64 v25, s11, 1.0
; SI-NEXT:    v_add_f32_e64 v24, s10, 1.0
; SI-NEXT:    v_lshr_b64 v[34:35], v[26:27], 16
; SI-NEXT:    v_lshr_b64 v[48:49], v[16:17], 16
; SI-NEXT:    v_add_f32_e64 v13, s29, 1.0
; SI-NEXT:    v_add_f32_e64 v12, s28, 1.0
; SI-NEXT:    v_add_f32_e64 v23, s13, 1.0
; SI-NEXT:    v_add_f32_e64 v22, s12, 1.0
; SI-NEXT:    v_lshr_b64 v[35:36], v[24:25], 16
; SI-NEXT:    v_lshr_b64 v[49:50], v[14:15], 16
; SI-NEXT:    v_add_f32_e64 v11, s27, 1.0
; SI-NEXT:    v_add_f32_e64 v10, s26, 1.0
; SI-NEXT:    v_add_f32_e64 v21, s15, 1.0
; SI-NEXT:    v_add_f32_e64 v20, s14, 1.0
; SI-NEXT:    v_lshr_b64 v[36:37], v[22:23], 16
; SI-NEXT:    v_lshr_b64 v[50:51], v[12:13], 16
; SI-NEXT:    v_add_f32_e64 v9, s25, 1.0
; SI-NEXT:    v_add_f32_e64 v8, s24, 1.0
; SI-NEXT:    v_add_f32_e64 v19, s41, 1.0
; SI-NEXT:    v_add_f32_e64 v18, s40, 1.0
; SI-NEXT:    v_lshr_b64 v[37:38], v[20:21], 16
; SI-NEXT:    v_lshr_b64 v[51:52], v[10:11], 16
; SI-NEXT:    v_add_f32_e64 v3, s19, 1.0
; SI-NEXT:    v_add_f32_e64 v2, s18, 1.0
; SI-NEXT:    v_add_f32_e64 v7, s23, 1.0
; SI-NEXT:    v_add_f32_e64 v6, s22, 1.0
; SI-NEXT:    v_lshr_b64 v[38:39], v[18:19], 16
; SI-NEXT:    v_lshr_b64 v[52:53], v[8:9], 16
; SI-NEXT:    v_add_f32_e64 v1, s17, 1.0
; SI-NEXT:    v_add_f32_e64 v0, s16, 1.0
; SI-NEXT:    v_add_f32_e64 v5, s21, 1.0
; SI-NEXT:    v_add_f32_e64 v4, s20, 1.0
; SI-NEXT:    v_lshr_b64 v[53:54], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[39:40], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[54:55], v[4:5], 16
; SI-NEXT:    v_lshr_b64 v[40:41], v[0:1], 16
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; SI-NEXT:    s_branch .LBB49_5
; SI-NEXT:  .LBB49_3:
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    s_branch .LBB49_2
; SI-NEXT:  .LBB49_4:
; SI-NEXT:    v_mov_b32_e32 v33, s30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_mov_b32_e32 v32, s38
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v62, s39
; SI-NEXT:    v_mov_b32_e32 v61, s48
; SI-NEXT:    v_mov_b32_e32 v60, s49
; SI-NEXT:    v_mov_b32_e32 v59, s50
; SI-NEXT:    v_mov_b32_e32 v58, s51
; SI-NEXT:    v_mov_b32_e32 v57, s52
; SI-NEXT:    v_mov_b32_e32 v56, s53
; SI-NEXT:    v_mov_b32_e32 v47, s54
; SI-NEXT:    v_mov_b32_e32 v46, s55
; SI-NEXT:    v_mov_b32_e32 v45, s64
; SI-NEXT:    v_mov_b32_e32 v44, s65
; SI-NEXT:    v_mov_b32_e32 v43, s66
; SI-NEXT:    v_mov_b32_e32 v42, s67
; SI-NEXT:    v_mov_b32_e32 v41, s68
; SI-NEXT:    v_mov_b32_e32 v55, s69
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v33, s34
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v34, s36
; SI-NEXT:    v_mov_b32_e32 v35, s46
; SI-NEXT:    v_mov_b32_e32 v36, s56
; SI-NEXT:    v_mov_b32_e32 v37, s58
; SI-NEXT:    v_mov_b32_e32 v38, s60
; SI-NEXT:    v_mov_b32_e32 v48, s62
; SI-NEXT:    v_mov_b32_e32 v49, s72
; SI-NEXT:    v_mov_b32_e32 v50, s74
; SI-NEXT:    v_mov_b32_e32 v51, s76
; SI-NEXT:    v_mov_b32_e32 v52, s78
; SI-NEXT:    v_mov_b32_e32 v53, s88
; SI-NEXT:    v_mov_b32_e32 v54, s90
; SI-NEXT:    v_mov_b32_e32 v39, s92
; SI-NEXT:    v_mov_b32_e32 v40, s94
; SI-NEXT:  .LBB49_5: ; %end
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v1, v1, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v39
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v32
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v62
; SI-NEXT:    v_or_b32_e32 v3, v3, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v54
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v32
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v61
; SI-NEXT:    v_or_b32_e32 v5, v5, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v53
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v60
; SI-NEXT:    v_or_b32_e32 v7, v7, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v52
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v32
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v59
; SI-NEXT:    v_or_b32_e32 v9, v9, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v32
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v58
; SI-NEXT:    v_or_b32_e32 v11, v11, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v32
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v57
; SI-NEXT:    v_or_b32_e32 v13, v13, v32
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v49
; SI-NEXT:    v_or_b32_e32 v14, v14, v32
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v56
; SI-NEXT:    v_or_b32_e32 v15, v15, v32
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v48
; SI-NEXT:    v_or_b32_e32 v16, v16, v32
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v47
; SI-NEXT:    v_or_b32_e32 v17, v17, v32
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v38
; SI-NEXT:    v_or_b32_e32 v18, v18, v32
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v46
; SI-NEXT:    v_or_b32_e32 v19, v19, v32
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v37
; SI-NEXT:    v_or_b32_e32 v20, v20, v32
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v45
; SI-NEXT:    v_or_b32_e32 v21, v21, v32
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v36
; SI-NEXT:    v_or_b32_e32 v22, v22, v32
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v44
; SI-NEXT:    v_or_b32_e32 v23, v23, v32
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v35
; SI-NEXT:    v_or_b32_e32 v24, v24, v32
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v43
; SI-NEXT:    v_or_b32_e32 v25, v25, v32
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v34
; SI-NEXT:    v_or_b32_e32 v26, v26, v32
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v27, v27, v32
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; SI-NEXT:    v_or_b32_e32 v0, v0, v40
; SI-NEXT:    v_or_b32_e32 v28, v28, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v41
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v55
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v32f32_to_v64i16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s36, 2
; VI-NEXT:    v_writelane_b32 v32, s37, 3
; VI-NEXT:    v_writelane_b32 v32, s38, 4
; VI-NEXT:    v_writelane_b32 v32, s39, 5
; VI-NEXT:    v_writelane_b32 v32, s48, 6
; VI-NEXT:    v_writelane_b32 v32, s49, 7
; VI-NEXT:    v_writelane_b32 v32, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB49_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB49_4
; VI-NEXT:  .LBB49_2: ; %cmp.true
; VI-NEXT:    v_add_f32_e64 v15, s31, 1.0
; VI-NEXT:    v_add_f32_e64 v14, s30, 1.0
; VI-NEXT:    v_add_f32_e64 v13, s29, 1.0
; VI-NEXT:    v_add_f32_e64 v12, s28, 1.0
; VI-NEXT:    v_add_f32_e64 v11, s27, 1.0
; VI-NEXT:    v_add_f32_e64 v10, s26, 1.0
; VI-NEXT:    v_add_f32_e64 v9, s25, 1.0
; VI-NEXT:    v_add_f32_e64 v8, s24, 1.0
; VI-NEXT:    v_add_f32_e64 v7, s23, 1.0
; VI-NEXT:    v_add_f32_e64 v6, s22, 1.0
; VI-NEXT:    v_add_f32_e64 v5, s21, 1.0
; VI-NEXT:    v_add_f32_e64 v4, s20, 1.0
; VI-NEXT:    v_add_f32_e64 v3, s19, 1.0
; VI-NEXT:    v_add_f32_e64 v2, s18, 1.0
; VI-NEXT:    v_add_f32_e64 v1, s17, 1.0
; VI-NEXT:    v_add_f32_e64 v0, s16, 1.0
; VI-NEXT:    v_add_f32_e64 v31, s51, 1.0
; VI-NEXT:    v_add_f32_e64 v30, s50, 1.0
; VI-NEXT:    v_add_f32_e64 v29, s49, 1.0
; VI-NEXT:    v_add_f32_e64 v28, s48, 1.0
; VI-NEXT:    v_add_f32_e64 v27, s47, 1.0
; VI-NEXT:    v_add_f32_e64 v26, s46, 1.0
; VI-NEXT:    v_add_f32_e64 v25, s45, 1.0
; VI-NEXT:    v_add_f32_e64 v24, s44, 1.0
; VI-NEXT:    v_add_f32_e64 v23, s43, 1.0
; VI-NEXT:    v_add_f32_e64 v22, s42, 1.0
; VI-NEXT:    v_add_f32_e64 v21, s41, 1.0
; VI-NEXT:    v_add_f32_e64 v20, s40, 1.0
; VI-NEXT:    v_add_f32_e64 v19, s39, 1.0
; VI-NEXT:    v_add_f32_e64 v18, s38, 1.0
; VI-NEXT:    v_add_f32_e64 v17, s37, 1.0
; VI-NEXT:    v_add_f32_e64 v16, s36, 1.0
; VI-NEXT:    s_branch .LBB49_5
; VI-NEXT:  .LBB49_3:
; VI-NEXT:    s_branch .LBB49_2
; VI-NEXT:  .LBB49_4:
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:  .LBB49_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v32, 9
; VI-NEXT:    v_readlane_b32 s50, v32, 8
; VI-NEXT:    v_readlane_b32 s49, v32, 7
; VI-NEXT:    v_readlane_b32 s48, v32, 6
; VI-NEXT:    v_readlane_b32 s39, v32, 5
; VI-NEXT:    v_readlane_b32 s38, v32, 4
; VI-NEXT:    v_readlane_b32 s37, v32, 3
; VI-NEXT:    v_readlane_b32 s36, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v32f32_to_v64i16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB49_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB49_4
; GFX9-NEXT:  .LBB49_2: ; %cmp.true
; GFX9-NEXT:    v_add_f32_e64 v15, s31, 1.0
; GFX9-NEXT:    v_add_f32_e64 v14, s30, 1.0
; GFX9-NEXT:    v_add_f32_e64 v13, s29, 1.0
; GFX9-NEXT:    v_add_f32_e64 v12, s28, 1.0
; GFX9-NEXT:    v_add_f32_e64 v11, s27, 1.0
; GFX9-NEXT:    v_add_f32_e64 v10, s26, 1.0
; GFX9-NEXT:    v_add_f32_e64 v9, s25, 1.0
; GFX9-NEXT:    v_add_f32_e64 v8, s24, 1.0
; GFX9-NEXT:    v_add_f32_e64 v7, s23, 1.0
; GFX9-NEXT:    v_add_f32_e64 v6, s22, 1.0
; GFX9-NEXT:    v_add_f32_e64 v5, s21, 1.0
; GFX9-NEXT:    v_add_f32_e64 v4, s20, 1.0
; GFX9-NEXT:    v_add_f32_e64 v3, s19, 1.0
; GFX9-NEXT:    v_add_f32_e64 v2, s18, 1.0
; GFX9-NEXT:    v_add_f32_e64 v1, s17, 1.0
; GFX9-NEXT:    v_add_f32_e64 v0, s16, 1.0
; GFX9-NEXT:    v_add_f32_e64 v31, s51, 1.0
; GFX9-NEXT:    v_add_f32_e64 v30, s50, 1.0
; GFX9-NEXT:    v_add_f32_e64 v29, s49, 1.0
; GFX9-NEXT:    v_add_f32_e64 v28, s48, 1.0
; GFX9-NEXT:    v_add_f32_e64 v27, s47, 1.0
; GFX9-NEXT:    v_add_f32_e64 v26, s46, 1.0
; GFX9-NEXT:    v_add_f32_e64 v25, s45, 1.0
; GFX9-NEXT:    v_add_f32_e64 v24, s44, 1.0
; GFX9-NEXT:    v_add_f32_e64 v23, s43, 1.0
; GFX9-NEXT:    v_add_f32_e64 v22, s42, 1.0
; GFX9-NEXT:    v_add_f32_e64 v21, s41, 1.0
; GFX9-NEXT:    v_add_f32_e64 v20, s40, 1.0
; GFX9-NEXT:    v_add_f32_e64 v19, s39, 1.0
; GFX9-NEXT:    v_add_f32_e64 v18, s38, 1.0
; GFX9-NEXT:    v_add_f32_e64 v17, s37, 1.0
; GFX9-NEXT:    v_add_f32_e64 v16, s36, 1.0
; GFX9-NEXT:    s_branch .LBB49_5
; GFX9-NEXT:  .LBB49_3:
; GFX9-NEXT:    s_branch .LBB49_2
; GFX9-NEXT:  .LBB49_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:  .LBB49_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v32f32_to_v64i16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB49_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB49_4
; GFX11-NEXT:  .LBB49_2: ; %cmp.true
; GFX11-NEXT:    v_add_f32_e64 v15, s27, 1.0
; GFX11-NEXT:    v_add_f32_e64 v14, s26, 1.0
; GFX11-NEXT:    v_add_f32_e64 v13, s25, 1.0
; GFX11-NEXT:    v_add_f32_e64 v12, s24, 1.0
; GFX11-NEXT:    v_add_f32_e64 v11, s23, 1.0
; GFX11-NEXT:    v_add_f32_e64 v10, s22, 1.0
; GFX11-NEXT:    v_add_f32_e64 v9, s21, 1.0
; GFX11-NEXT:    v_add_f32_e64 v8, s20, 1.0
; GFX11-NEXT:    v_add_f32_e64 v7, s19, 1.0
; GFX11-NEXT:    v_add_f32_e64 v6, s18, 1.0
; GFX11-NEXT:    v_add_f32_e64 v5, s17, 1.0
; GFX11-NEXT:    v_add_f32_e64 v4, s16, 1.0
; GFX11-NEXT:    v_add_f32_e64 v3, s15, 1.0
; GFX11-NEXT:    v_add_f32_e64 v2, s14, 1.0
; GFX11-NEXT:    v_add_f32_e64 v1, s13, 1.0
; GFX11-NEXT:    v_add_f32_e64 v0, s12, 1.0
; GFX11-NEXT:    v_add_f32_e64 v31, s51, 1.0
; GFX11-NEXT:    v_add_f32_e64 v30, s50, 1.0
; GFX11-NEXT:    v_add_f32_e64 v29, s49, 1.0
; GFX11-NEXT:    v_add_f32_e64 v28, s48, 1.0
; GFX11-NEXT:    v_add_f32_e64 v27, s47, 1.0
; GFX11-NEXT:    v_add_f32_e64 v26, s46, 1.0
; GFX11-NEXT:    v_add_f32_e64 v25, s45, 1.0
; GFX11-NEXT:    v_add_f32_e64 v24, s44, 1.0
; GFX11-NEXT:    v_add_f32_e64 v23, s43, 1.0
; GFX11-NEXT:    v_add_f32_e64 v22, s42, 1.0
; GFX11-NEXT:    v_add_f32_e64 v21, s41, 1.0
; GFX11-NEXT:    v_add_f32_e64 v20, s40, 1.0
; GFX11-NEXT:    v_add_f32_e64 v19, s39, 1.0
; GFX11-NEXT:    v_add_f32_e64 v18, s38, 1.0
; GFX11-NEXT:    v_add_f32_e64 v17, s37, 1.0
; GFX11-NEXT:    v_add_f32_e64 v16, s36, 1.0
; GFX11-NEXT:    s_branch .LBB49_5
; GFX11-NEXT:  .LBB49_3:
; GFX11-NEXT:    s_branch .LBB49_2
; GFX11-NEXT:  .LBB49_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:  .LBB49_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <32 x float> %a, splat (float 1.000000e+00)
  %a2 = bitcast <32 x float> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <32 x float> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define <32 x float> @bitcast_v64i16_to_v32f32(<64 x i16> %a, i32 %b) {
; SI-LABEL: bitcast_v64i16_to_v32f32:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v19
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v20
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v21
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v22
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v23
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v24
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v25
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v27
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v17
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v61
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v23
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v22
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v21
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v19
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v18
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v17
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v10
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v62
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v28
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v29
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v30
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v63
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB50_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_or_b32_e32 v8, v8, v32
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    v_or_b32_e32 v0, v0, v49
; SI-NEXT:    v_or_b32_e32 v1, v1, v42
; SI-NEXT:    v_or_b32_e32 v2, v2, v34
; SI-NEXT:    v_or_b32_e32 v3, v3, v41
; SI-NEXT:    v_or_b32_e32 v4, v4, v48
; SI-NEXT:    v_or_b32_e32 v5, v5, v40
; SI-NEXT:    v_or_b32_e32 v6, v6, v33
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v54
; SI-NEXT:    v_or_b32_e32 v10, v10, v39
; SI-NEXT:    v_or_b32_e32 v11, v11, v53
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v52
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v37
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v15, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v17, v50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v18, v35
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v63
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB50_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB50_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v46
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v32, v8
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v62
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v61
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v60
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v59
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v58
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v57
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v56
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v47
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v44
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    s_mov_b32 s6, 0x30000
; SI-NEXT:    v_or_b32_e32 v2, v34, v2
; SI-NEXT:    v_or_b32_e32 v3, v41, v3
; SI-NEXT:    v_or_b32_e32 v4, v48, v4
; SI-NEXT:    v_or_b32_e32 v5, v40, v5
; SI-NEXT:    v_or_b32_e32 v6, v33, v6
; SI-NEXT:    v_or_b32_e32 v7, v55, v7
; SI-NEXT:    v_or_b32_e32 v9, v54, v9
; SI-NEXT:    v_or_b32_e32 v10, v39, v10
; SI-NEXT:    v_or_b32_e32 v11, v53, v11
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v38, v12
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v52, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v37, v14
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v51, v15
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v36, v16
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v50, v17
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v35, v18
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
; SI-NEXT:  .LBB50_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v32f32:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB50_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v33, 3
; VI-NEXT:    v_add_u16_e32 v32, 3, v15
; VI-NEXT:    v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v32, v15
; VI-NEXT:    v_add_u16_e32 v32, 3, v14
; VI-NEXT:    v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v32, v14
; VI-NEXT:    v_add_u16_e32 v32, 3, v13
; VI-NEXT:    v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v32, v13
; VI-NEXT:    v_add_u16_e32 v32, 3, v12
; VI-NEXT:    v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v32, v12
; VI-NEXT:    v_add_u16_e32 v32, 3, v11
; VI-NEXT:    v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v32, v11
; VI-NEXT:    v_add_u16_e32 v32, 3, v10
; VI-NEXT:    v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v32, v10
; VI-NEXT:    v_add_u16_e32 v32, 3, v9
; VI-NEXT:    v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v32, v9
; VI-NEXT:    v_add_u16_e32 v32, 3, v8
; VI-NEXT:    v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v32, v8
; VI-NEXT:    v_add_u16_e32 v32, 3, v7
; VI-NEXT:    v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v32, v7
; VI-NEXT:    v_add_u16_e32 v32, 3, v6
; VI-NEXT:    v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v32, v6
; VI-NEXT:    v_add_u16_e32 v32, 3, v5
; VI-NEXT:    v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v32, v5
; VI-NEXT:    v_add_u16_e32 v32, 3, v4
; VI-NEXT:    v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v32, v4
; VI-NEXT:    v_add_u16_e32 v32, 3, v3
; VI-NEXT:    v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v32, v3
; VI-NEXT:    v_add_u16_e32 v32, 3, v2
; VI-NEXT:    v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v32, v2
; VI-NEXT:    v_add_u16_e32 v32, 3, v1
; VI-NEXT:    v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v32, v1
; VI-NEXT:    v_add_u16_e32 v32, 3, v0
; VI-NEXT:    v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v32, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v31
; VI-NEXT:    v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:    v_add_u16_e32 v32, 3, v30
; VI-NEXT:    v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v32, v30
; VI-NEXT:    v_add_u16_e32 v32, 3, v29
; VI-NEXT:    v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v32, v29
; VI-NEXT:    v_add_u16_e32 v32, 3, v28
; VI-NEXT:    v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v32, v28
; VI-NEXT:    v_add_u16_e32 v32, 3, v27
; VI-NEXT:    v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v32, v27
; VI-NEXT:    v_add_u16_e32 v32, 3, v26
; VI-NEXT:    v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v32, v26
; VI-NEXT:    v_add_u16_e32 v32, 3, v25
; VI-NEXT:    v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v32, v25
; VI-NEXT:    v_add_u16_e32 v32, 3, v24
; VI-NEXT:    v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v32, v24
; VI-NEXT:    v_add_u16_e32 v32, 3, v23
; VI-NEXT:    v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v32, v23
; VI-NEXT:    v_add_u16_e32 v32, 3, v22
; VI-NEXT:    v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v32, v22
; VI-NEXT:    v_add_u16_e32 v32, 3, v21
; VI-NEXT:    v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v32, v21
; VI-NEXT:    v_add_u16_e32 v32, 3, v20
; VI-NEXT:    v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v32, v20
; VI-NEXT:    v_add_u16_e32 v32, 3, v19
; VI-NEXT:    v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v32, v19
; VI-NEXT:    v_add_u16_e32 v32, 3, v18
; VI-NEXT:    v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v32, v18
; VI-NEXT:    v_add_u16_e32 v32, 3, v17
; VI-NEXT:    v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB50_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64i16_to_v32f32:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB50_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB50_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v32f32:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB50_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:  .LBB50_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define inreg <32 x float> @bitcast_v64i16_to_v32f32_scalar(<64 x i16> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64i16_to_v32f32_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    v_writelane_b32 v32, s70, 22
; SI-NEXT:    v_writelane_b32 v32, s71, 23
; SI-NEXT:    v_writelane_b32 v32, s80, 24
; SI-NEXT:    v_writelane_b32 v32, s81, 25
; SI-NEXT:    v_writelane_b32 v32, s82, 26
; SI-NEXT:    v_writelane_b32 v32, s83, 27
; SI-NEXT:    v_writelane_b32 v32, s84, 28
; SI-NEXT:    v_writelane_b32 v32, s85, 29
; SI-NEXT:    v_writelane_b32 v32, s86, 30
; SI-NEXT:    v_writelane_b32 v32, s87, 31
; SI-NEXT:    v_writelane_b32 v32, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s9, v16
; SI-NEXT:    v_writelane_b32 v32, s97, 33
; SI-NEXT:    s_lshr_b32 s14, s9, 16
; SI-NEXT:    v_readfirstlane_b32 s13, v14
; SI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v32, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s7, v17
; SI-NEXT:    v_readfirstlane_b32 s11, v15
; SI-NEXT:    s_lshr_b32 s72, s13, 16
; SI-NEXT:    v_readfirstlane_b32 s15, v13
; SI-NEXT:    v_readfirstlane_b32 s73, v12
; SI-NEXT:    v_readfirstlane_b32 s75, v11
; SI-NEXT:    v_readfirstlane_b32 s77, v10
; SI-NEXT:    v_readfirstlane_b32 s79, v9
; SI-NEXT:    v_readfirstlane_b32 s89, v8
; SI-NEXT:    v_readfirstlane_b32 s91, v7
; SI-NEXT:    v_readfirstlane_b32 s94, v6
; SI-NEXT:    v_readfirstlane_b32 s31, v5
; SI-NEXT:    v_readfirstlane_b32 s68, v4
; SI-NEXT:    v_readfirstlane_b32 s80, v3
; SI-NEXT:    v_readfirstlane_b32 s83, v2
; SI-NEXT:    v_readfirstlane_b32 s86, v1
; SI-NEXT:    v_readfirstlane_b32 s97, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v33, s14, 0
; SI-NEXT:    v_writelane_b32 v32, s99, 35
; SI-NEXT:    s_lshr_b32 s92, s29, 16
; SI-NEXT:    s_lshr_b32 s95, s28, 16
; SI-NEXT:    s_lshr_b32 s34, s27, 16
; SI-NEXT:    s_lshr_b32 s69, s26, 16
; SI-NEXT:    s_lshr_b32 s71, s25, 16
; SI-NEXT:    s_lshr_b32 s81, s24, 16
; SI-NEXT:    s_lshr_b32 s84, s23, 16
; SI-NEXT:    s_lshr_b32 s87, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s6, s19, 16
; SI-NEXT:    s_lshr_b32 s8, s18, 16
; SI-NEXT:    s_lshr_b32 s10, s17, 16
; SI-NEXT:    s_lshr_b32 s12, s16, 16
; SI-NEXT:    s_lshr_b32 vcc_lo, s7, 16
; SI-NEXT:    s_lshr_b32 vcc_hi, s11, 16
; SI-NEXT:    s_lshr_b32 s63, s15, 16
; SI-NEXT:    s_lshr_b32 s62, s73, 16
; SI-NEXT:    s_lshr_b32 s61, s75, 16
; SI-NEXT:    s_lshr_b32 s60, s77, 16
; SI-NEXT:    s_lshr_b32 s78, s79, 16
; SI-NEXT:    s_lshr_b32 s88, s89, 16
; SI-NEXT:    s_lshr_b32 s90, s91, 16
; SI-NEXT:    s_lshr_b32 s93, s94, 16
; SI-NEXT:    s_lshr_b32 s30, s31, 16
; SI-NEXT:    s_lshr_b32 s35, s68, 16
; SI-NEXT:    s_lshr_b32 s70, s80, 16
; SI-NEXT:    s_lshr_b32 s82, s83, 16
; SI-NEXT:    s_lshr_b32 s85, s86, 16
; SI-NEXT:    s_lshr_b32 s96, s97, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v33, s72, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v33, s63, 2
; SI-NEXT:    v_writelane_b32 v33, s62, 3
; SI-NEXT:    s_cbranch_scc0 .LBB51_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s10, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s60, 16
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s61, 16
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s62, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s63, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB51_3
; SI-NEXT:  .LBB51_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_or_b32 s5, s10, s5
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s36, s4, 0x30000
; SI-NEXT:    s_add_i32 s37, s5, 0x30000
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s38, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s39, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s40, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s41, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s42, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s43, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s44, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s45, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s46, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s47, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s48, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s97, s97, 3
; SI-NEXT:    s_add_i32 s49, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s86, s86, 3
; SI-NEXT:    s_add_i32 s50, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s83, s83, 3
; SI-NEXT:    s_add_i32 s51, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s80, s80, 3
; SI-NEXT:    s_add_i32 s52, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s68, s68, 3
; SI-NEXT:    s_add_i32 s53, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s31, s31, 3
; SI-NEXT:    s_add_i32 s54, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s94, s94, 3
; SI-NEXT:    s_add_i32 s55, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s91, s91, 3
; SI-NEXT:    s_add_i32 s56, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s89, s89, 3
; SI-NEXT:    s_add_i32 s57, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s79, s79, 3
; SI-NEXT:    s_add_i32 s58, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s77, s77, 3
; SI-NEXT:    s_add_i32 s59, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s76, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s75, s75, 3
; SI-NEXT:    s_add_i32 s60, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s74, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s73, s73, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 3
; SI-NEXT:    s_add_i32 s61, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 2
; SI-NEXT:    s_add_i32 s62, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 1
; SI-NEXT:    s_add_i32 s63, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s64, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 0
; SI-NEXT:    s_add_i32 s65, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s66, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s67, s4, 0x30000
; SI-NEXT:  .LBB51_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:    v_readlane_b32 s99, v32, 35
; SI-NEXT:    v_readlane_b32 s98, v32, 34
; SI-NEXT:    v_readlane_b32 s97, v32, 33
; SI-NEXT:    v_readlane_b32 s96, v32, 32
; SI-NEXT:    v_readlane_b32 s87, v32, 31
; SI-NEXT:    v_readlane_b32 s86, v32, 30
; SI-NEXT:    v_readlane_b32 s85, v32, 29
; SI-NEXT:    v_readlane_b32 s84, v32, 28
; SI-NEXT:    v_readlane_b32 s83, v32, 27
; SI-NEXT:    v_readlane_b32 s82, v32, 26
; SI-NEXT:    v_readlane_b32 s81, v32, 25
; SI-NEXT:    v_readlane_b32 s80, v32, 24
; SI-NEXT:    v_readlane_b32 s71, v32, 23
; SI-NEXT:    v_readlane_b32 s70, v32, 22
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB51_4:
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB51_2
;
; VI-LABEL: bitcast_v64i16_to_v32f32_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB51_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB51_3
; VI-NEXT:  .LBB51_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s5, s46, 3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s47, 3
; VI-NEXT:    s_add_i32 s46, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s29, 3
; VI-NEXT:    s_add_i32 s47, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s28, 3
; VI-NEXT:    s_add_i32 s29, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s28, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s27, 3
; VI-NEXT:    s_add_i32 s28, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s27, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s26, 3
; VI-NEXT:    s_add_i32 s27, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s26, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s25, 3
; VI-NEXT:    s_add_i32 s26, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s25, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s24, 3
; VI-NEXT:    s_add_i32 s25, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s24, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s23, 3
; VI-NEXT:    s_add_i32 s24, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s23, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s22, 3
; VI-NEXT:    s_add_i32 s23, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s22, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s21, 3
; VI-NEXT:    s_add_i32 s22, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s21, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s20, 3
; VI-NEXT:    s_add_i32 s21, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s20, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s19, 3
; VI-NEXT:    s_add_i32 s20, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s19, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s18, 3
; VI-NEXT:    s_add_i32 s19, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s18, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s17, 3
; VI-NEXT:    s_add_i32 s18, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s17, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s16, 3
; VI-NEXT:    s_add_i32 s17, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s16, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s6, 3
; VI-NEXT:    s_add_i32 s16, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s6, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s7, 3
; VI-NEXT:    s_add_i32 s6, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s7, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s8, 3
; VI-NEXT:    s_add_i32 s7, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s8, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s9, 3
; VI-NEXT:    s_add_i32 s8, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s9, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s10, 3
; VI-NEXT:    s_add_i32 s9, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s10, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s11, 3
; VI-NEXT:    s_add_i32 s10, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s11, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s12, 3
; VI-NEXT:    s_add_i32 s11, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s12, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s13, 3
; VI-NEXT:    s_add_i32 s12, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s13, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s14, 3
; VI-NEXT:    s_add_i32 s13, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s14, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s15, 3
; VI-NEXT:    s_add_i32 s14, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s15, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s40, 3
; VI-NEXT:    s_add_i32 s15, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s41, 3
; VI-NEXT:    s_add_i32 s40, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s42, 3
; VI-NEXT:    s_add_i32 s41, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s43, 3
; VI-NEXT:    s_add_i32 s42, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s44, 3
; VI-NEXT:    s_add_i32 s43, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s45, 3
; VI-NEXT:    s_add_i32 s44, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s45, s4, 0x30000
; VI-NEXT:  .LBB51_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB51_4:
; VI-NEXT:    s_branch .LBB51_2
;
; GFX9-LABEL: bitcast_v64i16_to_v32f32_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB51_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB51_4
; GFX9-NEXT:  .LBB51_2: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB51_5
; GFX9-NEXT:  .LBB51_3:
; GFX9-NEXT:    s_branch .LBB51_2
; GFX9-NEXT:  .LBB51_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB51_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v32f32_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB51_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB51_4
; GFX11-NEXT:  .LBB51_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_branch .LBB51_5
; GFX11-NEXT:  .LBB51_3:
; GFX11-NEXT:    s_branch .LBB51_2
; GFX11-NEXT:  .LBB51_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB51_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <32 x float>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <32 x float>
  br label %end

end:
  %phi = phi <32 x float> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <32 x float> %phi
}

define <16 x double> @bitcast_v16i64_to_v16f64(<16 x i64> %a, i32 %b) {
; SI-LABEL: bitcast_v16i64_to_v16f64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB52_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; SI-NEXT:  .LBB52_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16i64_to_v16f64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB52_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; VI-NEXT:  .LBB52_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16i64_to_v16f64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB52_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v0
; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v2
; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v4
; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v6
; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v8
; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v10
; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v12
; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v14
; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v16
; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v18, vcc, 3, v18
; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v20
; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v22
; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v24
; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v26
; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v28
; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
; GFX9-NEXT:  .LBB52_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16i64_to_v16f64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB52_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 3
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
; GFX11-NEXT:  .LBB52_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define inreg <16 x double> @bitcast_v16i64_to_v16f64_scalar(<16 x i64> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16i64_to_v16f64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s10, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s12, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s14, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s42, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s47, v0
; SI-NEXT:    s_cbranch_scc0 .LBB53_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB53_3
; SI-NEXT:  .LBB53_2: ; %cmp.true
; SI-NEXT:    s_add_u32 s16, s16, 3
; SI-NEXT:    s_addc_u32 s17, s17, 0
; SI-NEXT:    s_add_u32 s18, s18, 3
; SI-NEXT:    s_addc_u32 s19, s19, 0
; SI-NEXT:    s_add_u32 s20, s20, 3
; SI-NEXT:    s_addc_u32 s21, s21, 0
; SI-NEXT:    s_add_u32 s22, s22, 3
; SI-NEXT:    s_addc_u32 s23, s23, 0
; SI-NEXT:    s_add_u32 s24, s24, 3
; SI-NEXT:    s_addc_u32 s25, s25, 0
; SI-NEXT:    s_add_u32 s26, s26, 3
; SI-NEXT:    s_addc_u32 s27, s27, 0
; SI-NEXT:    s_add_u32 s28, s28, 3
; SI-NEXT:    s_addc_u32 s29, s29, 0
; SI-NEXT:    s_add_u32 s47, s47, 3
; SI-NEXT:    s_addc_u32 s46, s46, 0
; SI-NEXT:    s_add_u32 s45, s45, 3
; SI-NEXT:    s_addc_u32 s44, s44, 0
; SI-NEXT:    s_add_u32 s43, s43, 3
; SI-NEXT:    s_addc_u32 s42, s42, 0
; SI-NEXT:    s_add_u32 s41, s41, 3
; SI-NEXT:    s_addc_u32 s40, s40, 0
; SI-NEXT:    s_add_u32 s15, s15, 3
; SI-NEXT:    s_addc_u32 s14, s14, 0
; SI-NEXT:    s_add_u32 s13, s13, 3
; SI-NEXT:    s_addc_u32 s12, s12, 0
; SI-NEXT:    s_add_u32 s11, s11, 3
; SI-NEXT:    s_addc_u32 s10, s10, 0
; SI-NEXT:    s_add_u32 s9, s9, 3
; SI-NEXT:    s_addc_u32 s8, s8, 0
; SI-NEXT:    s_add_u32 s7, s7, 3
; SI-NEXT:    s_addc_u32 s6, s6, 0
; SI-NEXT:  .LBB53_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s47
; SI-NEXT:    v_mov_b32_e32 v15, s46
; SI-NEXT:    v_mov_b32_e32 v16, s45
; SI-NEXT:    v_mov_b32_e32 v17, s44
; SI-NEXT:    v_mov_b32_e32 v18, s43
; SI-NEXT:    v_mov_b32_e32 v19, s42
; SI-NEXT:    v_mov_b32_e32 v20, s41
; SI-NEXT:    v_mov_b32_e32 v21, s40
; SI-NEXT:    v_mov_b32_e32 v22, s15
; SI-NEXT:    v_mov_b32_e32 v23, s14
; SI-NEXT:    v_mov_b32_e32 v24, s13
; SI-NEXT:    v_mov_b32_e32 v25, s12
; SI-NEXT:    v_mov_b32_e32 v26, s11
; SI-NEXT:    v_mov_b32_e32 v27, s10
; SI-NEXT:    v_mov_b32_e32 v28, s9
; SI-NEXT:    v_mov_b32_e32 v29, s8
; SI-NEXT:    v_mov_b32_e32 v30, s7
; SI-NEXT:    v_mov_b32_e32 v31, s6
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB53_4:
; SI-NEXT:    s_branch .LBB53_2
;
; VI-LABEL: bitcast_v16i64_to_v16f64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB53_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB53_3
; VI-NEXT:  .LBB53_2: ; %cmp.true
; VI-NEXT:    s_add_u32 s16, s16, 3
; VI-NEXT:    s_addc_u32 s17, s17, 0
; VI-NEXT:    s_add_u32 s18, s18, 3
; VI-NEXT:    s_addc_u32 s19, s19, 0
; VI-NEXT:    s_add_u32 s20, s20, 3
; VI-NEXT:    s_addc_u32 s21, s21, 0
; VI-NEXT:    s_add_u32 s22, s22, 3
; VI-NEXT:    s_addc_u32 s23, s23, 0
; VI-NEXT:    s_add_u32 s24, s24, 3
; VI-NEXT:    s_addc_u32 s25, s25, 0
; VI-NEXT:    s_add_u32 s26, s26, 3
; VI-NEXT:    s_addc_u32 s27, s27, 0
; VI-NEXT:    s_add_u32 s28, s28, 3
; VI-NEXT:    s_addc_u32 s29, s29, 0
; VI-NEXT:    s_add_u32 s47, s47, 3
; VI-NEXT:    s_addc_u32 s46, s46, 0
; VI-NEXT:    s_add_u32 s45, s45, 3
; VI-NEXT:    s_addc_u32 s44, s44, 0
; VI-NEXT:    s_add_u32 s43, s43, 3
; VI-NEXT:    s_addc_u32 s42, s42, 0
; VI-NEXT:    s_add_u32 s41, s41, 3
; VI-NEXT:    s_addc_u32 s40, s40, 0
; VI-NEXT:    s_add_u32 s15, s15, 3
; VI-NEXT:    s_addc_u32 s14, s14, 0
; VI-NEXT:    s_add_u32 s13, s13, 3
; VI-NEXT:    s_addc_u32 s12, s12, 0
; VI-NEXT:    s_add_u32 s11, s11, 3
; VI-NEXT:    s_addc_u32 s10, s10, 0
; VI-NEXT:    s_add_u32 s9, s9, 3
; VI-NEXT:    s_addc_u32 s8, s8, 0
; VI-NEXT:    s_add_u32 s7, s7, 3
; VI-NEXT:    s_addc_u32 s6, s6, 0
; VI-NEXT:  .LBB53_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB53_4:
; VI-NEXT:    s_branch .LBB53_2
;
; GFX9-LABEL: bitcast_v16i64_to_v16f64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB53_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB53_3
; GFX9-NEXT:  .LBB53_2: ; %cmp.true
; GFX9-NEXT:    s_add_u32 s16, s16, 3
; GFX9-NEXT:    s_addc_u32 s17, s17, 0
; GFX9-NEXT:    s_add_u32 s18, s18, 3
; GFX9-NEXT:    s_addc_u32 s19, s19, 0
; GFX9-NEXT:    s_add_u32 s20, s20, 3
; GFX9-NEXT:    s_addc_u32 s21, s21, 0
; GFX9-NEXT:    s_add_u32 s22, s22, 3
; GFX9-NEXT:    s_addc_u32 s23, s23, 0
; GFX9-NEXT:    s_add_u32 s24, s24, 3
; GFX9-NEXT:    s_addc_u32 s25, s25, 0
; GFX9-NEXT:    s_add_u32 s26, s26, 3
; GFX9-NEXT:    s_addc_u32 s27, s27, 0
; GFX9-NEXT:    s_add_u32 s28, s28, 3
; GFX9-NEXT:    s_addc_u32 s29, s29, 0
; GFX9-NEXT:    s_add_u32 s47, s47, 3
; GFX9-NEXT:    s_addc_u32 s46, s46, 0
; GFX9-NEXT:    s_add_u32 s45, s45, 3
; GFX9-NEXT:    s_addc_u32 s44, s44, 0
; GFX9-NEXT:    s_add_u32 s43, s43, 3
; GFX9-NEXT:    s_addc_u32 s42, s42, 0
; GFX9-NEXT:    s_add_u32 s41, s41, 3
; GFX9-NEXT:    s_addc_u32 s40, s40, 0
; GFX9-NEXT:    s_add_u32 s15, s15, 3
; GFX9-NEXT:    s_addc_u32 s14, s14, 0
; GFX9-NEXT:    s_add_u32 s13, s13, 3
; GFX9-NEXT:    s_addc_u32 s12, s12, 0
; GFX9-NEXT:    s_add_u32 s11, s11, 3
; GFX9-NEXT:    s_addc_u32 s10, s10, 0
; GFX9-NEXT:    s_add_u32 s9, s9, 3
; GFX9-NEXT:    s_addc_u32 s8, s8, 0
; GFX9-NEXT:    s_add_u32 s7, s7, 3
; GFX9-NEXT:    s_addc_u32 s6, s6, 0
; GFX9-NEXT:  .LBB53_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB53_4:
; GFX9-NEXT:    s_branch .LBB53_2
;
; GFX11-LABEL: bitcast_v16i64_to_v16f64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB53_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB53_3
; GFX11-NEXT:  .LBB53_2: ; %cmp.true
; GFX11-NEXT:    s_add_u32 s0, s0, 3
; GFX11-NEXT:    s_addc_u32 s1, s1, 0
; GFX11-NEXT:    s_add_u32 s2, s2, 3
; GFX11-NEXT:    s_addc_u32 s3, s3, 0
; GFX11-NEXT:    s_add_u32 s16, s16, 3
; GFX11-NEXT:    s_addc_u32 s17, s17, 0
; GFX11-NEXT:    s_add_u32 s18, s18, 3
; GFX11-NEXT:    s_addc_u32 s19, s19, 0
; GFX11-NEXT:    s_add_u32 s20, s20, 3
; GFX11-NEXT:    s_addc_u32 s21, s21, 0
; GFX11-NEXT:    s_add_u32 s22, s22, 3
; GFX11-NEXT:    s_addc_u32 s23, s23, 0
; GFX11-NEXT:    s_add_u32 s24, s24, 3
; GFX11-NEXT:    s_addc_u32 s25, s25, 0
; GFX11-NEXT:    s_add_u32 s26, s26, 3
; GFX11-NEXT:    s_addc_u32 s27, s27, 0
; GFX11-NEXT:    s_add_u32 s28, s28, 3
; GFX11-NEXT:    s_addc_u32 s29, s29, 0
; GFX11-NEXT:    s_add_u32 s41, s41, 3
; GFX11-NEXT:    s_addc_u32 s40, s40, 0
; GFX11-NEXT:    s_add_u32 s15, s15, 3
; GFX11-NEXT:    s_addc_u32 s14, s14, 0
; GFX11-NEXT:    s_add_u32 s13, s13, 3
; GFX11-NEXT:    s_addc_u32 s12, s12, 0
; GFX11-NEXT:    s_add_u32 s11, s11, 3
; GFX11-NEXT:    s_addc_u32 s10, s10, 0
; GFX11-NEXT:    s_add_u32 s9, s9, 3
; GFX11-NEXT:    s_addc_u32 s8, s8, 0
; GFX11-NEXT:    s_add_u32 s7, s7, 3
; GFX11-NEXT:    s_addc_u32 s6, s6, 0
; GFX11-NEXT:    s_add_u32 s5, s5, 3
; GFX11-NEXT:    s_addc_u32 s4, s4, 0
; GFX11-NEXT:  .LBB53_3: ; %end
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB53_4:
; GFX11-NEXT:    s_branch .LBB53_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define <16 x i64> @bitcast_v16f64_to_v16i64(<16 x double> %a, i32 %b) {
; SI-LABEL: bitcast_v16f64_to_v16i64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB54_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; SI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; SI-NEXT:  .LBB54_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v16i64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB54_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; VI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; VI-NEXT:  .LBB54_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v16i64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB54_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX9-NEXT:  .LBB54_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v16i64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB54_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX11-NEXT:  .LBB54_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define inreg <16 x i64> @bitcast_v16f64_to_v16i64_scalar(<16 x double> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16f64_to_v16i64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s36, 0
; SI-NEXT:    v_writelane_b32 v32, s37, 1
; SI-NEXT:    v_writelane_b32 v32, s38, 2
; SI-NEXT:    v_writelane_b32 v32, s39, 3
; SI-NEXT:    v_writelane_b32 v32, s48, 4
; SI-NEXT:    v_writelane_b32 v32, s49, 5
; SI-NEXT:    v_writelane_b32 v32, s50, 6
; SI-NEXT:    v_writelane_b32 v32, s51, 7
; SI-NEXT:    v_writelane_b32 v32, s52, 8
; SI-NEXT:    v_writelane_b32 v32, s53, 9
; SI-NEXT:    v_writelane_b32 v32, s54, 10
; SI-NEXT:    v_writelane_b32 v32, s55, 11
; SI-NEXT:    v_writelane_b32 v32, s64, 12
; SI-NEXT:    v_writelane_b32 v32, s65, 13
; SI-NEXT:    v_writelane_b32 v32, s66, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v32, s67, 15
; SI-NEXT:    s_mov_b32 s49, s29
; SI-NEXT:    s_mov_b32 s48, s28
; SI-NEXT:    s_mov_b32 s47, s27
; SI-NEXT:    s_mov_b32 s46, s26
; SI-NEXT:    s_mov_b32 s45, s25
; SI-NEXT:    s_mov_b32 s44, s24
; SI-NEXT:    s_mov_b32 s43, s23
; SI-NEXT:    s_mov_b32 s42, s22
; SI-NEXT:    s_mov_b32 s41, s21
; SI-NEXT:    s_mov_b32 s40, s20
; SI-NEXT:    s_mov_b32 s39, s19
; SI-NEXT:    s_mov_b32 s38, s18
; SI-NEXT:    s_mov_b32 s37, s17
; SI-NEXT:    s_mov_b32 s36, s16
; SI-NEXT:    v_readfirstlane_b32 s67, v17
; SI-NEXT:    v_readfirstlane_b32 s66, v16
; SI-NEXT:    v_readfirstlane_b32 s65, v15
; SI-NEXT:    v_readfirstlane_b32 s64, v14
; SI-NEXT:    v_readfirstlane_b32 s63, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v12
; SI-NEXT:    v_readfirstlane_b32 s61, v11
; SI-NEXT:    v_readfirstlane_b32 s60, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s58, v8
; SI-NEXT:    v_readfirstlane_b32 s57, v7
; SI-NEXT:    v_readfirstlane_b32 s56, v6
; SI-NEXT:    v_readfirstlane_b32 s55, v5
; SI-NEXT:    v_readfirstlane_b32 s54, v4
; SI-NEXT:    v_readfirstlane_b32 s53, v3
; SI-NEXT:    v_readfirstlane_b32 s52, v2
; SI-NEXT:    v_readfirstlane_b32 s51, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s50, v0
; SI-NEXT:    s_cbranch_scc0 .LBB55_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB55_4
; SI-NEXT:  .LBB55_2: ; %cmp.true
; SI-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; SI-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; SI-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; SI-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; SI-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; SI-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; SI-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; SI-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; SI-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; SI-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; SI-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; SI-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; SI-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; SI-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; SI-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; SI-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; SI-NEXT:    s_branch .LBB55_5
; SI-NEXT:  .LBB55_3:
; SI-NEXT:    s_branch .LBB55_2
; SI-NEXT:  .LBB55_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB55_5: ; %end
; SI-NEXT:    v_readlane_b32 s67, v32, 15
; SI-NEXT:    v_readlane_b32 s66, v32, 14
; SI-NEXT:    v_readlane_b32 s65, v32, 13
; SI-NEXT:    v_readlane_b32 s64, v32, 12
; SI-NEXT:    v_readlane_b32 s55, v32, 11
; SI-NEXT:    v_readlane_b32 s54, v32, 10
; SI-NEXT:    v_readlane_b32 s53, v32, 9
; SI-NEXT:    v_readlane_b32 s52, v32, 8
; SI-NEXT:    v_readlane_b32 s51, v32, 7
; SI-NEXT:    v_readlane_b32 s50, v32, 6
; SI-NEXT:    v_readlane_b32 s49, v32, 5
; SI-NEXT:    v_readlane_b32 s48, v32, 4
; SI-NEXT:    v_readlane_b32 s39, v32, 3
; SI-NEXT:    v_readlane_b32 s38, v32, 2
; SI-NEXT:    v_readlane_b32 s37, v32, 1
; SI-NEXT:    v_readlane_b32 s36, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v16i64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s36, 0
; VI-NEXT:    v_writelane_b32 v32, s37, 1
; VI-NEXT:    v_writelane_b32 v32, s38, 2
; VI-NEXT:    v_writelane_b32 v32, s39, 3
; VI-NEXT:    v_writelane_b32 v32, s48, 4
; VI-NEXT:    v_writelane_b32 v32, s49, 5
; VI-NEXT:    v_writelane_b32 v32, s50, 6
; VI-NEXT:    v_writelane_b32 v32, s51, 7
; VI-NEXT:    v_writelane_b32 v32, s52, 8
; VI-NEXT:    v_writelane_b32 v32, s53, 9
; VI-NEXT:    v_writelane_b32 v32, s54, 10
; VI-NEXT:    v_writelane_b32 v32, s55, 11
; VI-NEXT:    v_writelane_b32 v32, s64, 12
; VI-NEXT:    v_writelane_b32 v32, s65, 13
; VI-NEXT:    v_writelane_b32 v32, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB55_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB55_4
; VI-NEXT:  .LBB55_2: ; %cmp.true
; VI-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; VI-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; VI-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; VI-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; VI-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; VI-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; VI-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; VI-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; VI-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; VI-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; VI-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; VI-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; VI-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; VI-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; VI-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; VI-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; VI-NEXT:    s_branch .LBB55_5
; VI-NEXT:  .LBB55_3:
; VI-NEXT:    s_branch .LBB55_2
; VI-NEXT:  .LBB55_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB55_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v32, 15
; VI-NEXT:    v_readlane_b32 s66, v32, 14
; VI-NEXT:    v_readlane_b32 s65, v32, 13
; VI-NEXT:    v_readlane_b32 s64, v32, 12
; VI-NEXT:    v_readlane_b32 s55, v32, 11
; VI-NEXT:    v_readlane_b32 s54, v32, 10
; VI-NEXT:    v_readlane_b32 s53, v32, 9
; VI-NEXT:    v_readlane_b32 s52, v32, 8
; VI-NEXT:    v_readlane_b32 s51, v32, 7
; VI-NEXT:    v_readlane_b32 s50, v32, 6
; VI-NEXT:    v_readlane_b32 s49, v32, 5
; VI-NEXT:    v_readlane_b32 s48, v32, 4
; VI-NEXT:    v_readlane_b32 s39, v32, 3
; VI-NEXT:    v_readlane_b32 s38, v32, 2
; VI-NEXT:    v_readlane_b32 s37, v32, 1
; VI-NEXT:    v_readlane_b32 s36, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v16i64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB55_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB55_4
; GFX9-NEXT:  .LBB55_2: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; GFX9-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; GFX9-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; GFX9-NEXT:    s_branch .LBB55_5
; GFX9-NEXT:  .LBB55_3:
; GFX9-NEXT:    s_branch .LBB55_2
; GFX9-NEXT:  .LBB55_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB55_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v16i64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB55_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB55_4
; GFX11-NEXT:  .LBB55_2: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[0:1], s[36:37], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], s[38:39], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], s[40:41], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], s[42:43], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], s[46:47], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], s[48:49], 1.0
; GFX11-NEXT:    v_add_f64 v[14:15], s[50:51], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], s[52:53], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], s[54:55], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], s[56:57], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], s[58:59], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], s[60:61], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], s[62:63], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], s[64:65], 1.0
; GFX11-NEXT:    v_add_f64 v[30:31], s[66:67], 1.0
; GFX11-NEXT:    s_branch .LBB55_5
; GFX11-NEXT:  .LBB55_3:
; GFX11-NEXT:    s_branch .LBB55_2
; GFX11-NEXT:  .LBB55_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB55_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) {
; SI-LABEL: bitcast_v16i64_to_v128i8:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB56_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v52, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v53, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB56_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB56_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_addc_u32_e32 v4, vcc, 0, v4, vcc
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_addc_u32_e32 v8, vcc, 0, v8, vcc
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_addc_u32_e32 v10, vcc, 0, v10, vcc
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v12, vcc
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v14, vcc
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_addc_u32_e32 v16, vcc, 0, v16, vcc
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v18, vcc
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_addc_u32_e32 v20, vcc, 0, v20, vcc
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_addc_u32_e32 v22, vcc, 0, v22, vcc
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_addc_u32_e32 v24, vcc, 0, v24, vcc
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_addc_u32_e32 v26, vcc, 0, v26, vcc
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_addc_u32_e32 v28, vcc, 0, v28, vcc
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_addc_u32_e32 v30, vcc, 0, v30, vcc
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v31, v32, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v38, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v48, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v51, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v52, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v55, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v41, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v43, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v45, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v31
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v53, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v40, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v42, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v46, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB56_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v45
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v1, v1, v45
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v43
; SI-NEXT:    v_lshlrev_b32_e32 v41, 24, v41
; SI-NEXT:    v_or_b32_e32 v41, v41, v43
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v41
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v34
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v33
; SI-NEXT:    v_or_b32_e32 v2, v33, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v55
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v52
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v51
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v61
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v48
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v38
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v37
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v60
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v58
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v7
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v57
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v47
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v9
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v46
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v44
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v42
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v11
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v40
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v53
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v13
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v39
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v15
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v17
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v27
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v28
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v29
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v30
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v31
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16i64_to_v128i8:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB56_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v58, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v56, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v1
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:  .LBB56_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB56_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v1, vcc, 3, v1
; VI-NEXT:    v_addc_u32_e32 v2, vcc, 0, v2, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, 3, v3
; VI-NEXT:    v_addc_u32_e32 v4, vcc, 0, v4, vcc
; VI-NEXT:    v_add_u32_e32 v5, vcc, 3, v5
; VI-NEXT:    v_addc_u32_e32 v6, vcc, 0, v6, vcc
; VI-NEXT:    v_add_u32_e32 v7, vcc, 3, v7
; VI-NEXT:    v_addc_u32_e32 v8, vcc, 0, v8, vcc
; VI-NEXT:    v_add_u32_e32 v9, vcc, 3, v9
; VI-NEXT:    v_addc_u32_e32 v10, vcc, 0, v10, vcc
; VI-NEXT:    v_add_u32_e32 v11, vcc, 3, v11
; VI-NEXT:    v_addc_u32_e32 v12, vcc, 0, v12, vcc
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v13
; VI-NEXT:    v_addc_u32_e32 v14, vcc, 0, v14, vcc
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v15
; VI-NEXT:    v_addc_u32_e32 v16, vcc, 0, v16, vcc
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v17
; VI-NEXT:    v_addc_u32_e32 v18, vcc, 0, v18, vcc
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_addc_u32_e32 v20, vcc, 0, v20, vcc
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_addc_u32_e32 v22, vcc, 0, v22, vcc
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_addc_u32_e32 v24, vcc, 0, v24, vcc
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_addc_u32_e32 v26, vcc, 0, v26, vcc
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_addc_u32_e32 v28, vcc, 0, v28, vcc
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_addc_u32_e32 v30, vcc, 0, v30, vcc
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_addc_u32_e32 v32, vcc, 0, v32, vcc
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v58, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v43, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v56, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v1
; VI-NEXT:  .LBB56_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v55, v57, v41, s4
; VI-NEXT:    v_perm_b32 v1, v1, v48, s4
; VI-NEXT:    v_lshlrev_b32_e32 v48, 16, v55
; VI-NEXT:    v_or_b32_e32 v1, v1, v48
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v34, v56, s4
; VI-NEXT:    v_perm_b32 v2, v2, v53, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v47, v40, s4
; VI-NEXT:    v_perm_b32 v2, v3, v38, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v46, v45, s4
; VI-NEXT:    v_perm_b32 v1, v4, v52, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v33, v39, s4
; VI-NEXT:    v_perm_b32 v2, v5, v37, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v44, v43, s4
; VI-NEXT:    v_perm_b32 v1, v6, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v50, v54, s4
; VI-NEXT:    v_perm_b32 v2, v7, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v62, v61, s4
; VI-NEXT:    v_perm_b32 v1, v8, v36, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v9, v49, s4
; VI-NEXT:    v_perm_b32 v1, v60, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v59, v58, s4
; VI-NEXT:    v_perm_b32 v1, v10, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v11, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v12, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v13, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v14, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v15, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v16, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v17, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v18, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v19, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v20, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v21, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v22, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v23, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v24, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v25, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v26, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v27, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v28, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v29, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v30, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v31, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v32, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v42, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16i64_to_v128i8:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; kill: killed $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(39)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB56_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(40)
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v7
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v1
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; GFX9-NEXT:  .LBB56_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB56_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    v_add_co_u32_e32 v1, vcc, 3, v1
; GFX9-NEXT:    v_addc_co_u32_e32 v2, vcc, 0, v2, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v3, vcc, 3, v3
; GFX9-NEXT:    v_addc_co_u32_e32 v4, vcc, 0, v4, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v5, vcc, 3, v5
; GFX9-NEXT:    v_addc_co_u32_e32 v6, vcc, 0, v6, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v7, vcc, 3, v7
; GFX9-NEXT:    v_addc_co_u32_e32 v8, vcc, 0, v8, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v9, vcc, 3, v9
; GFX9-NEXT:    v_addc_co_u32_e32 v10, vcc, 0, v10, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v11, vcc, 3, v11
; GFX9-NEXT:    v_addc_co_u32_e32 v12, vcc, 0, v12, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v13, vcc, 3, v13
; GFX9-NEXT:    v_addc_co_u32_e32 v14, vcc, 0, v14, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v15, vcc, 3, v15
; GFX9-NEXT:    v_addc_co_u32_e32 v16, vcc, 0, v16, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v17, vcc, 3, v17
; GFX9-NEXT:    v_addc_co_u32_e32 v18, vcc, 0, v18, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v19, vcc, 3, v19
; GFX9-NEXT:    v_addc_co_u32_e32 v20, vcc, 0, v20, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v21, vcc, 3, v21
; GFX9-NEXT:    v_addc_co_u32_e32 v22, vcc, 0, v22, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v23, vcc, 3, v23
; GFX9-NEXT:    v_addc_co_u32_e32 v24, vcc, 0, v24, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v25, vcc, 3, v25
; GFX9-NEXT:    v_addc_co_u32_e32 v26, vcc, 0, v26, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v27, vcc, 3, v27
; GFX9-NEXT:    v_addc_co_u32_e32 v28, vcc, 0, v28, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v29, vcc, 3, v29
; GFX9-NEXT:    v_addc_co_u32_e32 v30, vcc, 0, v30, vcc
; GFX9-NEXT:    s_waitcnt vmcnt(38)
; GFX9-NEXT:    v_add_co_u32_e32 v31, vcc, 3, v31
; GFX9-NEXT:    v_addc_co_u32_e32 v32, vcc, 0, v32, vcc
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v7
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v1
; GFX9-NEXT:  .LBB56_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v11, v11, v36, s4
; GFX9-NEXT:    v_perm_b32 v12, v12, v35, s4
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v3, v3, v34, s4
; GFX9-NEXT:    v_perm_b32 v34, v38, v39, s4
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v4, v4, v51, s4
; GFX9-NEXT:    v_perm_b32 v6, v6, v33, s4
; GFX9-NEXT:    v_perm_b32 v33, v50, v54, s4
; GFX9-NEXT:    v_perm_b32 v55, v58, v41, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v59, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, v2, v52, s4
; GFX9-NEXT:    v_perm_b32 v37, v37, v53, s4
; GFX9-NEXT:    v_perm_b32 v52, v57, v40, s4
; GFX9-NEXT:    v_perm_b32 v5, v5, v56, s4
; GFX9-NEXT:    v_perm_b32 v7, v7, v47, s4
; GFX9-NEXT:    v_perm_b32 v8, v8, v46, s4
; GFX9-NEXT:    v_perm_b32 v9, v9, v45, s4
; GFX9-NEXT:    v_perm_b32 v10, v10, v44, s4
; GFX9-NEXT:    v_perm_b32 v13, v13, v60, s4
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v35, v36, v35, s4
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v38, v63, v38, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v15, v15, v39, s4
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v14, v14, v36, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v16, v16, v39, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v36, v36, v50, s4
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v39, v39, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v17, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v50, v50, v53, s4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX9-NEXT:    v_perm_b32 v1, v49, v43, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshl_or_b32 v1, v52, 16, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    v_perm_b32 v1, v48, v62, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshl_or_b32 v1, v34, 16, v5
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v61, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v6
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshl_or_b32 v1, v33, 16, v7
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v8
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshl_or_b32 v1, v37, 16, v9
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v10
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshl_or_b32 v1, v38, 16, v11
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v12
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshl_or_b32 v1, v35, 16, v13
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v14
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshl_or_b32 v1, v36, 16, v15
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v16
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshl_or_b32 v1, v39, 16, v17
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v18
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v19, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v50, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v20, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v21, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v22, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v23, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v24, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v25, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v26, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v27, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v28, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v30, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v31, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v32, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v42, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v16i64_to_v128i8:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x2
; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB56_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB56_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB56_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_co_u32 v1, vcc_lo, v1, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v2, null, 0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v3, vcc_lo, v3, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v4, null, 0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v5, vcc_lo, v5, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v6, null, 0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v7, vcc_lo, v7, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v8, null, 0, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v9, vcc_lo, v9, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v10, null, 0, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v11, vcc_lo, v11, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v12, null, 0, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v13, vcc_lo, v13, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v14, null, 0, v14, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v15, vcc_lo, v15, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v16, null, 0, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v17, vcc_lo, v17, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v18, null, 0, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v19, vcc_lo, v19, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v20, null, 0, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v21, vcc_lo, v21, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v22, null, 0, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v25, vcc_lo, v25, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v26, null, 0, v26, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v27, vcc_lo, v27, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v28, null, 0, v28, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v29, vcc_lo, v29, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v30, null, 0, v30, vcc_lo
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_add_co_u32 v31, vcc_lo, v31, 3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v32, null, 0, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_co_u32 v23, vcc_lo, v23, 3
; GFX11-TRUE16-NEXT:    v_add_co_ci_u32_e64 v24, null, 0, v24, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB56_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v1.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, v3.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v66, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v3, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, v4.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v65, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v6.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v39, 16, v2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v5.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, v8.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v7.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v66, v66, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v147, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v54, v64, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v39, 16, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v9.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v64, v65, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v66, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v55, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v10.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v11.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v64, 16, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v54, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v52, v55, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v12.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v13.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v14.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v15.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v12, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v13, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v14, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v15, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v16.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v17.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v18.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v19.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v16, v117, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v17, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v18, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v19, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v20.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v21.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v22.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v23.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v20, v103, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v21, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v99, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v14, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v24.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v15, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v25.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v26.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v27.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v27, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v28.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v29.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v30.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v31.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v28, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v29, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v30, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v31, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v32, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v29
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v16i64_to_v128i8:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x2
; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:12
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB56_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:  .LBB56_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB56_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_co_u32 v1, vcc_lo, v1, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v2, null, 0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v3, vcc_lo, v3, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v4, null, 0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v5, vcc_lo, v5, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v6, null, 0, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v7, vcc_lo, v7, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v8, null, 0, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v9, vcc_lo, v9, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v10, null, 0, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v11, vcc_lo, v11, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v12, null, 0, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v13, vcc_lo, v13, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v14, null, 0, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v15, vcc_lo, v15, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v16, null, 0, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v17, vcc_lo, v17, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v18, null, 0, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v19, vcc_lo, v19, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v20, null, 0, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v21, vcc_lo, v21, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v22, null, 0, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v25, vcc_lo, v25, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v26, null, 0, v26, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v27, vcc_lo, v27, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v28, null, 0, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v29, vcc_lo, v29, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v30, null, 0, v30, vcc_lo
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_add_co_u32 v31, vcc_lo, v31, 3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v32, null, 0, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_co_u32 v23, vcc_lo, v23, 3
; GFX11-FAKE16-NEXT:    v_add_co_ci_u32_e64 v24, null, 0, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:  .LBB56_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v74, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v73, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v72, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v61, v65, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v59, v58, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v43, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v55, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v57, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v56, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v64, v46, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v41, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v182, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v53, v180, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v177, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v55, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v64, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v65, 16, v8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v54, 16, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v13, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v164, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v14, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v162, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v15, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v151, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v16, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v149, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v17, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v146, v49, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v18, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v144, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v19, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v133, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v20, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v131, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v21, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v128, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v118, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v115, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v24, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v113, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v25, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v102, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v26, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v97, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v28, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v87, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v29, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v84, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v30, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v31, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v71, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v69, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:88
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define inreg <128 x i8> @bitcast_v16i64_to_v128i8_scalar(<16 x i64> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16i64_to_v128i8_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_writelane_b32 v20, s30, 0
; SI-NEXT:    v_writelane_b32 v20, s31, 1
; SI-NEXT:    v_writelane_b32 v20, s34, 2
; SI-NEXT:    v_writelane_b32 v20, s35, 3
; SI-NEXT:    v_writelane_b32 v20, s36, 4
; SI-NEXT:    v_writelane_b32 v20, s37, 5
; SI-NEXT:    v_writelane_b32 v20, s38, 6
; SI-NEXT:    v_writelane_b32 v20, s39, 7
; SI-NEXT:    v_writelane_b32 v20, s48, 8
; SI-NEXT:    v_writelane_b32 v20, s49, 9
; SI-NEXT:    v_writelane_b32 v20, s50, 10
; SI-NEXT:    v_writelane_b32 v20, s51, 11
; SI-NEXT:    v_writelane_b32 v20, s52, 12
; SI-NEXT:    v_writelane_b32 v20, s53, 13
; SI-NEXT:    v_writelane_b32 v20, s54, 14
; SI-NEXT:    v_writelane_b32 v20, s55, 15
; SI-NEXT:    v_writelane_b32 v20, s64, 16
; SI-NEXT:    v_writelane_b32 v20, s65, 17
; SI-NEXT:    v_writelane_b32 v20, s66, 18
; SI-NEXT:    v_writelane_b32 v20, s67, 19
; SI-NEXT:    v_writelane_b32 v20, s68, 20
; SI-NEXT:    v_writelane_b32 v20, s69, 21
; SI-NEXT:    v_writelane_b32 v20, s70, 22
; SI-NEXT:    v_writelane_b32 v20, s71, 23
; SI-NEXT:    v_writelane_b32 v20, s80, 24
; SI-NEXT:    v_writelane_b32 v20, s81, 25
; SI-NEXT:    v_writelane_b32 v20, s82, 26
; SI-NEXT:    v_writelane_b32 v20, s83, 27
; SI-NEXT:    v_writelane_b32 v20, s84, 28
; SI-NEXT:    v_writelane_b32 v20, s85, 29
; SI-NEXT:    v_writelane_b32 v20, s86, 30
; SI-NEXT:    v_writelane_b32 v20, s87, 31
; SI-NEXT:    v_writelane_b32 v20, s96, 32
; SI-NEXT:    v_writelane_b32 v20, s97, 33
; SI-NEXT:    v_readfirstlane_b32 s44, v19
; SI-NEXT:    v_writelane_b32 v20, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s5, v18
; SI-NEXT:    v_readfirstlane_b32 s4, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s6, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s8, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s10, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s12, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s14, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s40, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s42, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v1
; SI-NEXT:    v_writelane_b32 v20, s99, 35
; SI-NEXT:    ; implicit-def: $vgpr22 : SGPR spill to VGPR lane
; SI-NEXT:    ; implicit-def: $vgpr21 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB57_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s46, s5, 24
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v22, s46, 40
; SI-NEXT:    s_lshr_b32 s46, s5, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 41
; SI-NEXT:    s_lshr_b32 s46, s5, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 42
; SI-NEXT:    s_lshr_b32 s46, s7, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 43
; SI-NEXT:    s_lshr_b32 s46, s7, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 44
; SI-NEXT:    s_lshr_b32 s46, s7, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 45
; SI-NEXT:    s_lshr_b32 s46, s9, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 46
; SI-NEXT:    s_lshr_b32 s46, s9, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 47
; SI-NEXT:    s_lshr_b32 s46, s9, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 48
; SI-NEXT:    s_lshr_b32 s46, s11, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 49
; SI-NEXT:    s_lshr_b32 s46, s11, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 50
; SI-NEXT:    s_lshr_b32 s46, s11, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 51
; SI-NEXT:    s_lshr_b32 s46, s13, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 52
; SI-NEXT:    s_lshr_b32 s46, s13, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 53
; SI-NEXT:    s_lshr_b32 s46, s13, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 54
; SI-NEXT:    s_lshr_b32 s46, s15, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 55
; SI-NEXT:    s_lshr_b32 s46, s15, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 56
; SI-NEXT:    s_lshr_b32 s46, s15, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 57
; SI-NEXT:    s_lshr_b32 s46, s41, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 58
; SI-NEXT:    s_lshr_b32 s46, s41, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 59
; SI-NEXT:    s_lshr_b32 s46, s41, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 60
; SI-NEXT:    s_lshr_b32 s46, s43, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 61
; SI-NEXT:    s_lshr_b32 s46, s43, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 62
; SI-NEXT:    s_lshr_b32 s46, s43, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 63
; SI-NEXT:    s_lshr_b32 s46, s45, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 0
; SI-NEXT:    s_lshr_b32 s46, s45, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 1
; SI-NEXT:    s_lshr_b32 s46, s45, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 2
; SI-NEXT:    s_lshr_b32 s46, s29, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 3
; SI-NEXT:    s_lshr_b32 s46, s29, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 4
; SI-NEXT:    s_lshr_b32 s46, s29, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 5
; SI-NEXT:    s_lshr_b32 s46, s27, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 6
; SI-NEXT:    s_lshr_b32 s46, s27, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 7
; SI-NEXT:    s_lshr_b32 s46, s27, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 8
; SI-NEXT:    s_lshr_b32 s46, s25, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 9
; SI-NEXT:    s_lshr_b32 s46, s25, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 10
; SI-NEXT:    s_lshr_b32 s46, s25, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 11
; SI-NEXT:    s_lshr_b32 s46, s23, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 12
; SI-NEXT:    s_lshr_b32 s46, s23, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 13
; SI-NEXT:    s_lshr_b32 s46, s23, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 14
; SI-NEXT:    s_lshr_b32 s46, s21, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 15
; SI-NEXT:    s_lshr_b32 s46, s21, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 16
; SI-NEXT:    s_lshr_b32 s46, s21, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 38
; SI-NEXT:    v_writelane_b32 v22, s47, 39
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 36
; SI-NEXT:    v_writelane_b32 v22, s47, 37
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 34
; SI-NEXT:    v_writelane_b32 v22, s47, 35
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 32
; SI-NEXT:    v_writelane_b32 v22, s47, 33
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 30
; SI-NEXT:    v_writelane_b32 v22, s47, 31
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 28
; SI-NEXT:    v_writelane_b32 v22, s47, 29
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 26
; SI-NEXT:    v_writelane_b32 v22, s47, 27
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 24
; SI-NEXT:    v_writelane_b32 v22, s47, 25
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 22
; SI-NEXT:    v_writelane_b32 v22, s47, 23
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 20
; SI-NEXT:    v_writelane_b32 v22, s47, 21
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 18
; SI-NEXT:    v_writelane_b32 v22, s47, 19
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 16
; SI-NEXT:    v_writelane_b32 v22, s47, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 14
; SI-NEXT:    v_writelane_b32 v22, s47, 15
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 12
; SI-NEXT:    v_writelane_b32 v22, s47, 13
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 10
; SI-NEXT:    v_writelane_b32 v22, s47, 11
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 8
; SI-NEXT:    v_writelane_b32 v22, s47, 9
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 6
; SI-NEXT:    v_writelane_b32 v22, s47, 7
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 4
; SI-NEXT:    v_writelane_b32 v22, s47, 5
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 2
; SI-NEXT:    v_writelane_b32 v22, s47, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 0
; SI-NEXT:    s_lshr_b32 s49, s19, 24
; SI-NEXT:    s_lshr_b32 s48, s19, 16
; SI-NEXT:    s_lshr_b32 s50, s19, 8
; SI-NEXT:    s_lshr_b32 s51, s17, 24
; SI-NEXT:    s_lshr_b32 s52, s17, 16
; SI-NEXT:    s_lshr_b32 s53, s17, 8
; SI-NEXT:    s_lshr_b64 s[54:55], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v22, s47, 1
; SI-NEXT:    s_lshr_b64 s[64:65], s[42:43], 24
; SI-NEXT:    s_lshr_b64 s[66:67], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[68:69], s[42:43], 8
; SI-NEXT:    s_lshr_b64 s[70:71], s[44:45], 24
; SI-NEXT:    s_lshr_b64 s[80:81], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[84:85], s[28:29], 24
; SI-NEXT:    s_lshr_b64 s[86:87], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[96:97], s[28:29], 8
; SI-NEXT:    s_lshr_b64 s[98:99], s[26:27], 24
; SI-NEXT:    s_lshr_b64 s[46:47], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[26:27], 8
; SI-NEXT:    s_lshr_b64 s[58:59], s[24:25], 24
; SI-NEXT:    s_lshr_b64 s[60:61], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[24:25], 8
; SI-NEXT:    s_lshr_b64 s[72:73], s[22:23], 24
; SI-NEXT:    s_lshr_b64 s[74:75], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[22:23], 8
; SI-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; SI-NEXT:    s_lshr_b64 s[88:89], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[20:21], 8
; SI-NEXT:    s_lshr_b64 s[92:93], s[18:19], 24
; SI-NEXT:    s_lshr_b64 s[94:95], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[18:19], 8
; SI-NEXT:    s_lshr_b64 s[34:35], s[16:17], 24
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_lshr_b64 s[38:39], s[16:17], 8
; SI-NEXT:    s_cbranch_execnz .LBB57_3
; SI-NEXT:  .LBB57_2: ; %cmp.true
; SI-NEXT:    s_add_u32 s4, s4, 3
; SI-NEXT:    s_addc_u32 s5, s5, 0
; SI-NEXT:    s_add_u32 s6, s6, 3
; SI-NEXT:    s_addc_u32 s7, s7, 0
; SI-NEXT:    s_add_u32 s8, s8, 3
; SI-NEXT:    s_addc_u32 s9, s9, 0
; SI-NEXT:    s_add_u32 s10, s10, 3
; SI-NEXT:    s_addc_u32 s11, s11, 0
; SI-NEXT:    s_add_u32 s12, s12, 3
; SI-NEXT:    s_addc_u32 s13, s13, 0
; SI-NEXT:    s_add_u32 s14, s14, 3
; SI-NEXT:    s_addc_u32 s15, s15, 0
; SI-NEXT:    s_add_u32 s40, s40, 3
; SI-NEXT:    s_addc_u32 s41, s41, 0
; SI-NEXT:    s_add_u32 s42, s42, 3
; SI-NEXT:    s_addc_u32 s43, s43, 0
; SI-NEXT:    s_add_u32 s44, s44, 3
; SI-NEXT:    s_addc_u32 s45, s45, 0
; SI-NEXT:    s_add_u32 s28, s28, 3
; SI-NEXT:    s_addc_u32 s29, s29, 0
; SI-NEXT:    s_add_u32 s26, s26, 3
; SI-NEXT:    s_addc_u32 s27, s27, 0
; SI-NEXT:    s_add_u32 s24, s24, 3
; SI-NEXT:    s_addc_u32 s25, s25, 0
; SI-NEXT:    s_add_u32 s22, s22, 3
; SI-NEXT:    s_addc_u32 s23, s23, 0
; SI-NEXT:    s_add_u32 s20, s20, 3
; SI-NEXT:    s_addc_u32 s21, s21, 0
; SI-NEXT:    s_add_u32 s18, s18, 3
; SI-NEXT:    s_addc_u32 s19, s19, 0
; SI-NEXT:    s_add_u32 s16, s16, 3
; SI-NEXT:    s_addc_u32 s17, s17, 0
; SI-NEXT:    s_lshr_b32 s46, s5, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 40
; SI-NEXT:    s_lshr_b32 s46, s5, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 41
; SI-NEXT:    s_lshr_b32 s46, s5, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 42
; SI-NEXT:    s_lshr_b32 s46, s7, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 43
; SI-NEXT:    s_lshr_b32 s46, s7, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 44
; SI-NEXT:    s_lshr_b32 s46, s7, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 45
; SI-NEXT:    s_lshr_b32 s46, s9, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 46
; SI-NEXT:    s_lshr_b32 s46, s9, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 47
; SI-NEXT:    s_lshr_b32 s46, s9, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 48
; SI-NEXT:    s_lshr_b32 s46, s11, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 49
; SI-NEXT:    s_lshr_b32 s46, s11, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 50
; SI-NEXT:    s_lshr_b32 s46, s11, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 51
; SI-NEXT:    s_lshr_b32 s46, s13, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 52
; SI-NEXT:    s_lshr_b32 s46, s13, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 53
; SI-NEXT:    s_lshr_b32 s46, s13, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 54
; SI-NEXT:    s_lshr_b32 s46, s15, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 55
; SI-NEXT:    s_lshr_b32 s46, s15, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 56
; SI-NEXT:    s_lshr_b32 s46, s15, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 57
; SI-NEXT:    s_lshr_b32 s46, s41, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 58
; SI-NEXT:    s_lshr_b32 s46, s41, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 59
; SI-NEXT:    s_lshr_b32 s46, s41, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 60
; SI-NEXT:    s_lshr_b32 s46, s43, 24
; SI-NEXT:    v_writelane_b32 v22, s46, 61
; SI-NEXT:    s_lshr_b32 s46, s43, 16
; SI-NEXT:    v_writelane_b32 v22, s46, 62
; SI-NEXT:    s_lshr_b32 s46, s43, 8
; SI-NEXT:    v_writelane_b32 v22, s46, 63
; SI-NEXT:    s_lshr_b32 s46, s45, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 0
; SI-NEXT:    s_lshr_b32 s46, s45, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 1
; SI-NEXT:    s_lshr_b32 s46, s45, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 2
; SI-NEXT:    s_lshr_b32 s46, s29, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 3
; SI-NEXT:    s_lshr_b32 s46, s29, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 4
; SI-NEXT:    s_lshr_b32 s46, s29, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 5
; SI-NEXT:    s_lshr_b32 s46, s27, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 6
; SI-NEXT:    s_lshr_b32 s46, s27, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 7
; SI-NEXT:    s_lshr_b32 s46, s27, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 8
; SI-NEXT:    s_lshr_b32 s46, s25, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 9
; SI-NEXT:    s_lshr_b32 s46, s25, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 10
; SI-NEXT:    s_lshr_b32 s46, s25, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 11
; SI-NEXT:    s_lshr_b32 s46, s23, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 12
; SI-NEXT:    s_lshr_b32 s46, s23, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 13
; SI-NEXT:    s_lshr_b32 s46, s23, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 14
; SI-NEXT:    s_lshr_b32 s46, s21, 24
; SI-NEXT:    v_writelane_b32 v21, s46, 15
; SI-NEXT:    s_lshr_b32 s46, s21, 16
; SI-NEXT:    v_writelane_b32 v21, s46, 16
; SI-NEXT:    s_lshr_b32 s46, s21, 8
; SI-NEXT:    v_writelane_b32 v21, s46, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 38
; SI-NEXT:    v_writelane_b32 v22, s47, 39
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 36
; SI-NEXT:    v_writelane_b32 v22, s47, 37
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 34
; SI-NEXT:    v_writelane_b32 v22, s47, 35
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 32
; SI-NEXT:    v_writelane_b32 v22, s47, 33
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 30
; SI-NEXT:    v_writelane_b32 v22, s47, 31
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 28
; SI-NEXT:    v_writelane_b32 v22, s47, 29
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 26
; SI-NEXT:    v_writelane_b32 v22, s47, 27
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 24
; SI-NEXT:    v_writelane_b32 v22, s47, 25
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 22
; SI-NEXT:    v_writelane_b32 v22, s47, 23
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 20
; SI-NEXT:    v_writelane_b32 v22, s47, 21
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 18
; SI-NEXT:    v_writelane_b32 v22, s47, 19
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 16
; SI-NEXT:    v_writelane_b32 v22, s47, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 14
; SI-NEXT:    v_writelane_b32 v22, s47, 15
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 12
; SI-NEXT:    v_writelane_b32 v22, s47, 13
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 10
; SI-NEXT:    v_writelane_b32 v22, s47, 11
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 8
; SI-NEXT:    v_writelane_b32 v22, s47, 9
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 6
; SI-NEXT:    v_writelane_b32 v22, s47, 7
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 24
; SI-NEXT:    v_writelane_b32 v22, s46, 4
; SI-NEXT:    v_writelane_b32 v22, s47, 5
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v22, s46, 2
; SI-NEXT:    v_writelane_b32 v22, s47, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[40:41], 8
; SI-NEXT:    v_writelane_b32 v22, s46, 0
; SI-NEXT:    s_lshr_b32 s49, s19, 24
; SI-NEXT:    s_lshr_b32 s48, s19, 16
; SI-NEXT:    s_lshr_b32 s50, s19, 8
; SI-NEXT:    s_lshr_b32 s51, s17, 24
; SI-NEXT:    s_lshr_b32 s52, s17, 16
; SI-NEXT:    s_lshr_b32 s53, s17, 8
; SI-NEXT:    s_lshr_b64 s[54:55], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v22, s47, 1
; SI-NEXT:    s_lshr_b64 s[64:65], s[42:43], 24
; SI-NEXT:    s_lshr_b64 s[66:67], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[68:69], s[42:43], 8
; SI-NEXT:    s_lshr_b64 s[70:71], s[44:45], 24
; SI-NEXT:    s_lshr_b64 s[80:81], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[84:85], s[28:29], 24
; SI-NEXT:    s_lshr_b64 s[86:87], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[96:97], s[28:29], 8
; SI-NEXT:    s_lshr_b64 s[98:99], s[26:27], 24
; SI-NEXT:    s_lshr_b64 s[46:47], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[26:27], 8
; SI-NEXT:    s_lshr_b64 s[58:59], s[24:25], 24
; SI-NEXT:    s_lshr_b64 s[60:61], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[24:25], 8
; SI-NEXT:    s_lshr_b64 s[72:73], s[22:23], 24
; SI-NEXT:    s_lshr_b64 s[74:75], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[22:23], 8
; SI-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; SI-NEXT:    s_lshr_b64 s[88:89], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[20:21], 8
; SI-NEXT:    s_lshr_b64 s[92:93], s[18:19], 24
; SI-NEXT:    s_lshr_b64 s[94:95], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[18:19], 8
; SI-NEXT:    s_lshr_b64 s[34:35], s[16:17], 24
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_lshr_b64 s[38:39], s[16:17], 8
; SI-NEXT:  .LBB57_3: ; %end
; SI-NEXT:    s_lshl_b32 s47, s38, 8
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    s_and_b32 s47, s36, 0xff
; SI-NEXT:    s_lshl_b32 s57, s34, 24
; SI-NEXT:    s_lshl_b32 s47, s47, 16
; SI-NEXT:    s_or_b32 s47, s57, s47
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    v_mov_b32_e32 v1, s16
; SI-NEXT:    s_and_b32 s16, s17, 0xff
; SI-NEXT:    s_lshl_b32 s17, s53, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s52, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s47, s51, 24
; SI-NEXT:    s_or_b32 s17, s47, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_lshl_b32 s16, s30, 8
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s94, 0xff
; SI-NEXT:    s_lshl_b32 s18, s92, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v3, s16
; SI-NEXT:    s_and_b32 s16, s19, 0xff
; SI-NEXT:    s_lshl_b32 s17, s50, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s48, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s49, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v4, s16
; SI-NEXT:    s_lshl_b32 s16, s90, 8
; SI-NEXT:    s_and_b32 s17, s20, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s88, 0xff
; SI-NEXT:    s_lshl_b32 s18, s78, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 17
; SI-NEXT:    v_mov_b32_e32 v5, s16
; SI-NEXT:    s_and_b32 s16, s21, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 16
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 15
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v6, s16
; SI-NEXT:    s_lshl_b32 s16, s76, 8
; SI-NEXT:    s_and_b32 s17, s22, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s74, 0xff
; SI-NEXT:    s_lshl_b32 s18, s72, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 14
; SI-NEXT:    v_mov_b32_e32 v7, s16
; SI-NEXT:    s_and_b32 s16, s23, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 13
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 12
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v8, s16
; SI-NEXT:    s_lshl_b32 s16, s62, 8
; SI-NEXT:    s_and_b32 s17, s24, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s60, 0xff
; SI-NEXT:    s_lshl_b32 s18, s58, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 11
; SI-NEXT:    v_mov_b32_e32 v9, s16
; SI-NEXT:    s_and_b32 s16, s25, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 10
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 9
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v10, s16
; SI-NEXT:    s_lshl_b32 s16, s56, 8
; SI-NEXT:    s_and_b32 s17, s26, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s46, 0xff
; SI-NEXT:    s_lshl_b32 s18, s98, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 8
; SI-NEXT:    v_mov_b32_e32 v11, s16
; SI-NEXT:    s_and_b32 s16, s27, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 7
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 6
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v12, s16
; SI-NEXT:    s_lshl_b32 s16, s96, 8
; SI-NEXT:    s_and_b32 s17, s28, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s86, 0xff
; SI-NEXT:    s_lshl_b32 s18, s84, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 5
; SI-NEXT:    v_mov_b32_e32 v13, s16
; SI-NEXT:    s_and_b32 s16, s29, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 4
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 3
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v14, s16
; SI-NEXT:    s_lshl_b32 s16, s82, 8
; SI-NEXT:    s_and_b32 s17, s44, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s80, 0xff
; SI-NEXT:    s_lshl_b32 s18, s70, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 2
; SI-NEXT:    v_mov_b32_e32 v15, s16
; SI-NEXT:    s_and_b32 s16, s45, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v21, 1
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 0
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 4, v0
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 12, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v4, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 16, v0
; SI-NEXT:    v_mov_b32_e32 v16, s16
; SI-NEXT:    s_lshl_b32 s16, s68, 8
; SI-NEXT:    s_and_b32 s17, s42, 0xff
; SI-NEXT:    buffer_store_dword v5, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 20, v0
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s66, 0xff
; SI-NEXT:    buffer_store_dword v6, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 24, v0
; SI-NEXT:    s_lshl_b32 s18, s64, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    buffer_store_dword v7, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 28, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    buffer_store_dword v8, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 36, v0
; SI-NEXT:    v_readlane_b32 s17, v22, 63
; SI-NEXT:    buffer_store_dword v10, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 40, v0
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s43, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    buffer_store_dword v11, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 44, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 62
; SI-NEXT:    buffer_store_dword v12, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 48, v0
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 61
; SI-NEXT:    buffer_store_dword v13, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 52, v0
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    buffer_store_dword v14, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 56, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_readlane_b32 s18, v22, 0
; SI-NEXT:    buffer_store_dword v15, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 60, v0
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s19, v22, 1
; SI-NEXT:    buffer_store_dword v16, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 64, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_lshl_b32 s17, s18, 8
; SI-NEXT:    v_readlane_b32 s18, v22, 2
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s40, 0xff
; SI-NEXT:    v_readlane_b32 s19, v22, 3
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 4
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x44, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 60
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s41, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 59
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 58
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x48, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    v_readlane_b32 s16, v22, 6
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s17, v22, 7
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    v_readlane_b32 s19, v22, 5
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    v_readlane_b32 s16, v22, 8
; SI-NEXT:    v_readlane_b32 s17, v22, 9
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 10
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s17, s18, 24
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x4c, v0
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s14
; SI-NEXT:    s_and_b32 s14, s15, 0xff
; SI-NEXT:    v_readlane_b32 s15, v22, 57
; SI-NEXT:    s_lshl_b32 s15, s15, 8
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    v_readlane_b32 s15, v22, 56
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    v_readlane_b32 s16, v22, 55
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_lshl_b32 s16, s16, 24
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x50, v0
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s14
; SI-NEXT:    v_readlane_b32 s14, v22, 12
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s15, v22, 13
; SI-NEXT:    s_lshl_b32 s14, s14, 8
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    v_readlane_b32 s14, v22, 14
; SI-NEXT:    v_readlane_b32 s15, v22, 15
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s16, v22, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_lshl_b32 s15, s16, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x54, v0
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s12
; SI-NEXT:    s_and_b32 s12, s13, 0xff
; SI-NEXT:    v_readlane_b32 s13, v22, 54
; SI-NEXT:    s_lshl_b32 s13, s13, 8
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    v_readlane_b32 s13, v22, 53
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    v_readlane_b32 s14, v22, 52
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x58, v0
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s12
; SI-NEXT:    v_readlane_b32 s12, v22, 18
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s13, v22, 19
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    v_readlane_b32 s12, v22, 20
; SI-NEXT:    v_readlane_b32 s13, v22, 21
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s14, v22, 22
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s13, s14, 24
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x5c, v0
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s10
; SI-NEXT:    s_and_b32 s10, s11, 0xff
; SI-NEXT:    v_readlane_b32 s11, v22, 51
; SI-NEXT:    s_lshl_b32 s11, s11, 8
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    v_readlane_b32 s11, v22, 50
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    v_readlane_b32 s12, v22, 49
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_lshl_b32 s12, s12, 24
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x60, v0
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s10
; SI-NEXT:    v_readlane_b32 s10, v22, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s11, v22, 25
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    v_readlane_b32 s10, v22, 26
; SI-NEXT:    v_readlane_b32 s11, v22, 27
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s12, v22, 28
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s11, s12, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x64, v0
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    s_and_b32 s8, s9, 0xff
; SI-NEXT:    v_readlane_b32 s9, v22, 48
; SI-NEXT:    s_lshl_b32 s9, s9, 8
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    v_readlane_b32 s9, v22, 47
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    v_readlane_b32 s10, v22, 46
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x68, v0
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    v_readlane_b32 s8, v22, 30
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s9, v22, 31
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    v_readlane_b32 s8, v22, 32
; SI-NEXT:    v_readlane_b32 s9, v22, 33
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v22, 34
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s10, 24
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x6c, v0
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    s_and_b32 s6, s7, 0xff
; SI-NEXT:    v_readlane_b32 s7, v22, 45
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    v_readlane_b32 s7, v22, 44
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    v_readlane_b32 s8, v22, 43
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x70, v0
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_readlane_b32 s6, v22, 36
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    v_readlane_b32 s7, v22, 37
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s54, 0xff
; SI-NEXT:    v_readlane_b32 s8, v22, 38
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s8, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x74, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    s_and_b32 s4, s5, 0xff
; SI-NEXT:    v_readlane_b32 s5, v22, 42
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_readlane_b32 s5, v22, 41
; SI-NEXT:    s_and_b32 s5, s5, 0xff
; SI-NEXT:    v_readlane_b32 s6, v22, 40
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x78, v0
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    v_mov_b32_e32 v1, s4
; SI-NEXT:    v_readlane_b32 s19, v22, 11
; SI-NEXT:    v_readlane_b32 s17, v22, 17
; SI-NEXT:    v_readlane_b32 s15, v22, 23
; SI-NEXT:    v_readlane_b32 s13, v22, 29
; SI-NEXT:    v_readlane_b32 s11, v22, 35
; SI-NEXT:    v_readlane_b32 s9, v22, 39
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    v_readlane_b32 s99, v20, 35
; SI-NEXT:    v_readlane_b32 s98, v20, 34
; SI-NEXT:    v_readlane_b32 s97, v20, 33
; SI-NEXT:    v_readlane_b32 s96, v20, 32
; SI-NEXT:    v_readlane_b32 s87, v20, 31
; SI-NEXT:    v_readlane_b32 s86, v20, 30
; SI-NEXT:    v_readlane_b32 s85, v20, 29
; SI-NEXT:    v_readlane_b32 s84, v20, 28
; SI-NEXT:    v_readlane_b32 s83, v20, 27
; SI-NEXT:    v_readlane_b32 s82, v20, 26
; SI-NEXT:    v_readlane_b32 s81, v20, 25
; SI-NEXT:    v_readlane_b32 s80, v20, 24
; SI-NEXT:    v_readlane_b32 s71, v20, 23
; SI-NEXT:    v_readlane_b32 s70, v20, 22
; SI-NEXT:    v_readlane_b32 s69, v20, 21
; SI-NEXT:    v_readlane_b32 s68, v20, 20
; SI-NEXT:    v_readlane_b32 s67, v20, 19
; SI-NEXT:    v_readlane_b32 s66, v20, 18
; SI-NEXT:    v_readlane_b32 s65, v20, 17
; SI-NEXT:    v_readlane_b32 s64, v20, 16
; SI-NEXT:    v_readlane_b32 s55, v20, 15
; SI-NEXT:    v_readlane_b32 s54, v20, 14
; SI-NEXT:    v_readlane_b32 s53, v20, 13
; SI-NEXT:    v_readlane_b32 s52, v20, 12
; SI-NEXT:    v_readlane_b32 s51, v20, 11
; SI-NEXT:    v_readlane_b32 s50, v20, 10
; SI-NEXT:    v_readlane_b32 s49, v20, 9
; SI-NEXT:    v_readlane_b32 s48, v20, 8
; SI-NEXT:    v_readlane_b32 s39, v20, 7
; SI-NEXT:    v_readlane_b32 s38, v20, 6
; SI-NEXT:    v_readlane_b32 s37, v20, 5
; SI-NEXT:    v_readlane_b32 s36, v20, 4
; SI-NEXT:    v_readlane_b32 s35, v20, 3
; SI-NEXT:    v_readlane_b32 s34, v20, 2
; SI-NEXT:    v_readlane_b32 s31, v20, 1
; SI-NEXT:    v_readlane_b32 s30, v20, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB57_4:
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v22, s54, 0
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s55, 1
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 2
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s55, 3
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 4
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s55, 5
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 6
; SI-NEXT:    v_writelane_b32 v22, s55, 7
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 8
; SI-NEXT:    v_writelane_b32 v22, s55, 9
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 10
; SI-NEXT:    v_writelane_b32 v22, s55, 11
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 12
; SI-NEXT:    v_writelane_b32 v22, s55, 13
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 14
; SI-NEXT:    v_writelane_b32 v22, s55, 15
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 16
; SI-NEXT:    v_writelane_b32 v22, s55, 17
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 18
; SI-NEXT:    v_writelane_b32 v22, s55, 19
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 20
; SI-NEXT:    v_writelane_b32 v22, s55, 21
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 22
; SI-NEXT:    v_writelane_b32 v22, s55, 23
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 24
; SI-NEXT:    v_writelane_b32 v22, s55, 25
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 26
; SI-NEXT:    v_writelane_b32 v22, s55, 27
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 28
; SI-NEXT:    v_writelane_b32 v22, s55, 29
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 30
; SI-NEXT:    v_writelane_b32 v22, s55, 31
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 32
; SI-NEXT:    v_writelane_b32 v22, s55, 33
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 34
; SI-NEXT:    v_writelane_b32 v22, s55, 35
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 36
; SI-NEXT:    v_writelane_b32 v22, s55, 37
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v22, s54, 38
; SI-NEXT:    v_writelane_b32 v22, s55, 39
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    s_branch .LBB57_2
;
; VI-LABEL: bitcast_v16i64_to_v128i8_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s34, 2
; VI-NEXT:    v_writelane_b32 v32, s35, 3
; VI-NEXT:    v_writelane_b32 v32, s36, 4
; VI-NEXT:    v_writelane_b32 v32, s37, 5
; VI-NEXT:    v_writelane_b32 v32, s38, 6
; VI-NEXT:    v_writelane_b32 v32, s39, 7
; VI-NEXT:    v_writelane_b32 v32, s48, 8
; VI-NEXT:    v_writelane_b32 v32, s49, 9
; VI-NEXT:    v_writelane_b32 v32, s50, 10
; VI-NEXT:    v_writelane_b32 v32, s51, 11
; VI-NEXT:    v_writelane_b32 v32, s52, 12
; VI-NEXT:    v_writelane_b32 v32, s53, 13
; VI-NEXT:    v_writelane_b32 v32, s54, 14
; VI-NEXT:    v_writelane_b32 v32, s55, 15
; VI-NEXT:    v_writelane_b32 v32, s64, 16
; VI-NEXT:    v_writelane_b32 v32, s65, 17
; VI-NEXT:    v_writelane_b32 v32, s66, 18
; VI-NEXT:    v_writelane_b32 v32, s67, 19
; VI-NEXT:    v_writelane_b32 v32, s68, 20
; VI-NEXT:    v_writelane_b32 v32, s69, 21
; VI-NEXT:    v_writelane_b32 v32, s70, 22
; VI-NEXT:    v_writelane_b32 v32, s71, 23
; VI-NEXT:    v_writelane_b32 v32, s80, 24
; VI-NEXT:    v_writelane_b32 v32, s81, 25
; VI-NEXT:    v_writelane_b32 v32, s82, 26
; VI-NEXT:    v_writelane_b32 v32, s83, 27
; VI-NEXT:    v_writelane_b32 v32, s84, 28
; VI-NEXT:    v_writelane_b32 v32, s85, 29
; VI-NEXT:    v_readfirstlane_b32 s44, v19
; VI-NEXT:    v_writelane_b32 v32, s86, 30
; VI-NEXT:    v_readfirstlane_b32 s5, v18
; VI-NEXT:    v_readfirstlane_b32 s4, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s6, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s8, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s10, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s12, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s14, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s40, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s42, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    s_cmp_lg_u32 s44, 0
; VI-NEXT:    v_readfirstlane_b32 s44, v1
; VI-NEXT:    v_writelane_b32 v32, s87, 31
; VI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; VI-NEXT:    s_cbranch_scc0 .LBB57_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s14, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s14, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s41, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s41, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s40, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s40, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s43, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s43, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s42, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s42, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 57
; VI-NEXT:    s_lshr_b64 s[56:57], s[4:5], 24
; VI-NEXT:    v_writelane_b32 v33, s56, 0
; VI-NEXT:    s_lshr_b32 s51, s29, 8
; VI-NEXT:    s_lshr_b32 s86, s28, 8
; VI-NEXT:    s_lshr_b32 s52, s27, 8
; VI-NEXT:    s_lshr_b32 s66, s26, 8
; VI-NEXT:    s_lshr_b32 s64, s25, 8
; VI-NEXT:    s_lshr_b32 s50, s24, 8
; VI-NEXT:    s_lshr_b32 s71, s23, 24
; VI-NEXT:    s_lshr_b32 s70, s23, 16
; VI-NEXT:    s_lshr_b32 s55, s23, 8
; VI-NEXT:    s_lshr_b32 s65, s22, 8
; VI-NEXT:    s_lshr_b32 s68, s21, 24
; VI-NEXT:    s_lshr_b32 s54, s21, 8
; VI-NEXT:    s_lshr_b32 s84, s20, 16
; VI-NEXT:    s_lshr_b32 s69, s20, 8
; VI-NEXT:    s_lshr_b32 s80, s19, 24
; VI-NEXT:    s_lshr_b32 s87, s19, 16
; VI-NEXT:    s_lshr_b32 s46, s19, 8
; VI-NEXT:    s_lshr_b32 s81, s18, 16
; VI-NEXT:    s_lshr_b32 s47, s18, 8
; VI-NEXT:    s_lshr_b32 s85, s17, 24
; VI-NEXT:    s_lshr_b32 s83, s17, 16
; VI-NEXT:    s_lshr_b32 s67, s17, 8
; VI-NEXT:    s_lshr_b32 s53, s16, 16
; VI-NEXT:    s_lshr_b32 s82, s16, 8
; VI-NEXT:    v_writelane_b32 v33, s57, 1
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:    s_cbranch_execnz .LBB57_3
; VI-NEXT:  .LBB57_2: ; %cmp.true
; VI-NEXT:    s_add_u32 s16, s16, 3
; VI-NEXT:    s_addc_u32 s17, s17, 0
; VI-NEXT:    s_add_u32 s18, s18, 3
; VI-NEXT:    s_addc_u32 s19, s19, 0
; VI-NEXT:    s_add_u32 s20, s20, 3
; VI-NEXT:    s_addc_u32 s21, s21, 0
; VI-NEXT:    s_add_u32 s22, s22, 3
; VI-NEXT:    s_addc_u32 s23, s23, 0
; VI-NEXT:    s_add_u32 s24, s24, 3
; VI-NEXT:    s_addc_u32 s25, s25, 0
; VI-NEXT:    s_add_u32 s26, s26, 3
; VI-NEXT:    s_addc_u32 s27, s27, 0
; VI-NEXT:    s_add_u32 s28, s28, 3
; VI-NEXT:    s_addc_u32 s29, s29, 0
; VI-NEXT:    s_add_u32 s44, s44, 3
; VI-NEXT:    s_addc_u32 s45, s45, 0
; VI-NEXT:    s_add_u32 s42, s42, 3
; VI-NEXT:    s_addc_u32 s43, s43, 0
; VI-NEXT:    s_add_u32 s40, s40, 3
; VI-NEXT:    s_addc_u32 s41, s41, 0
; VI-NEXT:    s_add_u32 s14, s14, 3
; VI-NEXT:    s_addc_u32 s15, s15, 0
; VI-NEXT:    s_add_u32 s12, s12, 3
; VI-NEXT:    s_addc_u32 s13, s13, 0
; VI-NEXT:    s_add_u32 s10, s10, 3
; VI-NEXT:    s_addc_u32 s11, s11, 0
; VI-NEXT:    s_add_u32 s8, s8, 3
; VI-NEXT:    s_addc_u32 s9, s9, 0
; VI-NEXT:    s_add_u32 s6, s6, 3
; VI-NEXT:    s_addc_u32 s7, s7, 0
; VI-NEXT:    s_add_u32 s4, s4, 3
; VI-NEXT:    s_addc_u32 s5, s5, 0
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s14, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s14, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s41, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s41, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s40, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s40, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s43, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s43, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s42, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s42, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 57
; VI-NEXT:    s_lshr_b64 s[56:57], s[4:5], 24
; VI-NEXT:    v_writelane_b32 v33, s56, 0
; VI-NEXT:    s_lshr_b32 s51, s29, 8
; VI-NEXT:    s_lshr_b32 s86, s28, 8
; VI-NEXT:    s_lshr_b32 s52, s27, 8
; VI-NEXT:    s_lshr_b32 s66, s26, 8
; VI-NEXT:    s_lshr_b32 s64, s25, 8
; VI-NEXT:    s_lshr_b32 s50, s24, 8
; VI-NEXT:    s_lshr_b32 s71, s23, 24
; VI-NEXT:    s_lshr_b32 s70, s23, 16
; VI-NEXT:    s_lshr_b32 s55, s23, 8
; VI-NEXT:    s_lshr_b32 s65, s22, 8
; VI-NEXT:    s_lshr_b32 s68, s21, 24
; VI-NEXT:    s_lshr_b32 s54, s21, 8
; VI-NEXT:    s_lshr_b32 s84, s20, 16
; VI-NEXT:    s_lshr_b32 s69, s20, 8
; VI-NEXT:    s_lshr_b32 s80, s19, 24
; VI-NEXT:    s_lshr_b32 s87, s19, 16
; VI-NEXT:    s_lshr_b32 s46, s19, 8
; VI-NEXT:    s_lshr_b32 s81, s18, 16
; VI-NEXT:    s_lshr_b32 s47, s18, 8
; VI-NEXT:    s_lshr_b32 s85, s17, 24
; VI-NEXT:    s_lshr_b32 s83, s17, 16
; VI-NEXT:    s_lshr_b32 s67, s17, 8
; VI-NEXT:    s_lshr_b32 s53, s16, 16
; VI-NEXT:    s_lshr_b32 s82, s16, 8
; VI-NEXT:    v_writelane_b32 v33, s57, 1
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:  .LBB57_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v8, s48
; VI-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; VI-NEXT:    v_mov_b32_e32 v2, s82
; VI-NEXT:    v_perm_b32 v8, s53, v8, v1
; VI-NEXT:    v_perm_b32 v3, s16, v2, v1
; VI-NEXT:    v_mov_b32_e32 v26, s85
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v4, s67
; VI-NEXT:    v_perm_b32 v26, s83, v26, v1
; VI-NEXT:    v_or_b32_e32 v3, v3, v8
; VI-NEXT:    v_perm_b32 v2, s17, v4, v1
; VI-NEXT:    v_mov_b32_e32 v12, s38
; VI-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v26
; VI-NEXT:    v_mov_b32_e32 v4, s47
; VI-NEXT:    v_perm_b32 v12, s81, v12, v1
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 4, v0
; VI-NEXT:    v_perm_b32 v5, s18, v4, v1
; VI-NEXT:    v_mov_b32_e32 v27, s80
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v12
; VI-NEXT:    v_mov_b32_e32 v6, s46
; VI-NEXT:    v_perm_b32 v27, s87, v27, v1
; VI-NEXT:    v_or_b32_e32 v2, v5, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 8, v0
; VI-NEXT:    v_perm_b32 v4, s19, v6, v1
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_readlane_b32 s16, v33, 46
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
; VI-NEXT:    v_mov_b32_e32 v6, s69
; VI-NEXT:    v_mov_b32_e32 v24, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 44
; VI-NEXT:    v_perm_b32 v16, s84, v16, v1
; VI-NEXT:    v_or_b32_e32 v2, v4, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 12, v0
; VI-NEXT:    v_perm_b32 v7, s20, v6, v1
; VI-NEXT:    v_mov_b32_e32 v25, s16
; VI-NEXT:    v_mov_b32_e32 v28, s68
; VI-NEXT:    v_readlane_b32 s16, v33, 57
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v16
; VI-NEXT:    v_mov_b32_e32 v9, s54
; VI-NEXT:    v_perm_b32 v28, s16, v28, v1
; VI-NEXT:    v_or_b32_e32 v2, v7, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 16, v0
; VI-NEXT:    v_perm_b32 v6, s21, v9, v1
; VI-NEXT:    v_mov_b32_e32 v19, s34
; VI-NEXT:    v_readlane_b32 s16, v33, 56
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
; VI-NEXT:    v_mov_b32_e32 v9, s65
; VI-NEXT:    v_perm_b32 v19, s16, v19, v1
; VI-NEXT:    v_or_b32_e32 v2, v6, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 20, v0
; VI-NEXT:    v_perm_b32 v10, s22, v9, v1
; VI-NEXT:    v_mov_b32_e32 v29, s71
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v19
; VI-NEXT:    v_mov_b32_e32 v11, s55
; VI-NEXT:    v_perm_b32 v29, s70, v29, v1
; VI-NEXT:    v_or_b32_e32 v2, v10, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 24, v0
; VI-NEXT:    v_perm_b32 v9, s23, v11, v1
; VI-NEXT:    v_mov_b32_e32 v20, s30
; VI-NEXT:    v_readlane_b32 s16, v33, 55
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v29
; VI-NEXT:    v_mov_b32_e32 v11, s50
; VI-NEXT:    v_perm_b32 v20, s16, v20, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 53
; VI-NEXT:    v_or_b32_e32 v2, v9, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 28, v0
; VI-NEXT:    v_perm_b32 v13, s24, v11, v1
; VI-NEXT:    v_mov_b32_e32 v30, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 54
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v20
; VI-NEXT:    v_mov_b32_e32 v14, s64
; VI-NEXT:    v_perm_b32 v30, s16, v30, v1
; VI-NEXT:    v_or_b32_e32 v2, v13, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 32, v0
; VI-NEXT:    v_perm_b32 v11, s25, v14, v1
; VI-NEXT:    v_mov_b32_e32 v21, s90
; VI-NEXT:    v_readlane_b32 s16, v33, 52
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
; VI-NEXT:    v_mov_b32_e32 v14, s66
; VI-NEXT:    v_perm_b32 v21, s16, v21, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 50
; VI-NEXT:    v_or_b32_e32 v2, v11, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 36, v0
; VI-NEXT:    v_perm_b32 v15, s26, v14, v1
; VI-NEXT:    v_mov_b32_e32 v31, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 51
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v21
; VI-NEXT:    v_mov_b32_e32 v17, s52
; VI-NEXT:    v_perm_b32 v31, s16, v31, v1
; VI-NEXT:    v_or_b32_e32 v2, v15, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 40, v0
; VI-NEXT:    v_perm_b32 v14, s27, v17, v1
; VI-NEXT:    v_mov_b32_e32 v22, s88
; VI-NEXT:    v_readlane_b32 s16, v33, 49
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v31
; VI-NEXT:    v_mov_b32_e32 v17, s86
; VI-NEXT:    v_perm_b32 v22, s16, v22, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 47
; VI-NEXT:    v_or_b32_e32 v2, v14, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 44, v0
; VI-NEXT:    v_perm_b32 v18, s28, v17, v1
; VI-NEXT:    v_mov_b32_e32 v8, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 48
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v22
; VI-NEXT:    v_mov_b32_e32 v23, s51
; VI-NEXT:    v_perm_b32 v3, s16, v8, v1
; VI-NEXT:    v_or_b32_e32 v2, v18, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 48, v0
; VI-NEXT:    v_perm_b32 v17, s29, v23, v1
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; VI-NEXT:    v_mov_b32_e32 v23, s78
; VI-NEXT:    v_or_b32_e32 v2, v17, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 52, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 45
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, s16, v23, v1
; VI-NEXT:    v_perm_b32 v24, s44, v24, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v2, v24, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 56, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 42
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 43
; VI-NEXT:    v_perm_b32 v2, s16, v2, v1
; VI-NEXT:    v_perm_b32 v25, s45, v25, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v2, v25, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 60, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 41
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_mov_b32_e32 v3, s76
; VI-NEXT:    v_readlane_b32 s16, v33, 40
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s42, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 64, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 39
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 37
; VI-NEXT:    v_mov_b32_e32 v3, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 38
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s43, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x44, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 36
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_mov_b32_e32 v3, s74
; VI-NEXT:    v_readlane_b32 s16, v33, 35
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s40, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x48, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 34
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 32
; VI-NEXT:    v_mov_b32_e32 v3, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 33
; VI-NEXT:    v_perm_b32 v3, s16, v3, v1
; VI-NEXT:    v_perm_b32 v2, s41, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x4c, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 31
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_perm_b32 v2, s14, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s72
; VI-NEXT:    v_readlane_b32 s14, v33, 30
; VI-NEXT:    v_perm_b32 v3, s14, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x50, v0
; VI-NEXT:    v_readlane_b32 s14, v33, 29
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s14
; VI-NEXT:    v_readlane_b32 s14, v33, 27
; VI-NEXT:    v_mov_b32_e32 v3, s14
; VI-NEXT:    v_readlane_b32 s14, v33, 28
; VI-NEXT:    v_perm_b32 v3, s14, v3, v1
; VI-NEXT:    v_perm_b32 v2, s15, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x54, v0
; VI-NEXT:    v_readlane_b32 s14, v33, 26
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s14
; VI-NEXT:    v_perm_b32 v2, s12, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s62
; VI-NEXT:    v_readlane_b32 s12, v33, 25
; VI-NEXT:    v_perm_b32 v3, s12, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x58, v0
; VI-NEXT:    v_readlane_b32 s12, v33, 24
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s12
; VI-NEXT:    v_readlane_b32 s12, v33, 22
; VI-NEXT:    v_mov_b32_e32 v3, s12
; VI-NEXT:    v_readlane_b32 s12, v33, 23
; VI-NEXT:    v_perm_b32 v3, s12, v3, v1
; VI-NEXT:    v_perm_b32 v2, s13, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x5c, v0
; VI-NEXT:    v_readlane_b32 s12, v33, 21
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s12
; VI-NEXT:    v_perm_b32 v2, s10, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s60
; VI-NEXT:    v_readlane_b32 s10, v33, 20
; VI-NEXT:    v_perm_b32 v3, s10, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x60, v0
; VI-NEXT:    v_readlane_b32 s10, v33, 19
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s10
; VI-NEXT:    v_readlane_b32 s10, v33, 17
; VI-NEXT:    v_mov_b32_e32 v3, s10
; VI-NEXT:    v_readlane_b32 s10, v33, 18
; VI-NEXT:    v_perm_b32 v3, s10, v3, v1
; VI-NEXT:    v_perm_b32 v2, s11, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x64, v0
; VI-NEXT:    v_readlane_b32 s10, v33, 16
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s10
; VI-NEXT:    v_perm_b32 v2, s8, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s58
; VI-NEXT:    v_readlane_b32 s8, v33, 15
; VI-NEXT:    v_perm_b32 v3, s8, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x68, v0
; VI-NEXT:    v_readlane_b32 s8, v33, 14
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s8
; VI-NEXT:    v_readlane_b32 s8, v33, 12
; VI-NEXT:    v_mov_b32_e32 v3, s8
; VI-NEXT:    v_readlane_b32 s8, v33, 13
; VI-NEXT:    v_perm_b32 v3, s8, v3, v1
; VI-NEXT:    v_perm_b32 v2, s9, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x6c, v0
; VI-NEXT:    v_readlane_b32 s8, v33, 11
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s8
; VI-NEXT:    v_perm_b32 v2, s6, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s56
; VI-NEXT:    v_readlane_b32 s6, v33, 10
; VI-NEXT:    v_perm_b32 v3, s6, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x70, v0
; VI-NEXT:    v_readlane_b32 s6, v33, 9
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 7
; VI-NEXT:    v_mov_b32_e32 v3, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 8
; VI-NEXT:    v_perm_b32 v3, s6, v3, v1
; VI-NEXT:    v_perm_b32 v2, s7, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x74, v0
; VI-NEXT:    v_readlane_b32 s6, v33, 6
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 0
; VI-NEXT:    v_perm_b32 v2, s4, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s6
; VI-NEXT:    v_readlane_b32 s4, v33, 5
; VI-NEXT:    v_perm_b32 v3, s4, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x78, v0
; VI-NEXT:    v_readlane_b32 s4, v33, 4
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s4
; VI-NEXT:    v_readlane_b32 s4, v33, 2
; VI-NEXT:    v_mov_b32_e32 v3, s4
; VI-NEXT:    v_readlane_b32 s4, v33, 3
; VI-NEXT:    v_perm_b32 v2, s5, v2, v1
; VI-NEXT:    v_perm_b32 v1, s4, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    v_readlane_b32 s7, v33, 1
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_readlane_b32 s87, v32, 31
; VI-NEXT:    v_readlane_b32 s86, v32, 30
; VI-NEXT:    v_readlane_b32 s85, v32, 29
; VI-NEXT:    v_readlane_b32 s84, v32, 28
; VI-NEXT:    v_readlane_b32 s83, v32, 27
; VI-NEXT:    v_readlane_b32 s82, v32, 26
; VI-NEXT:    v_readlane_b32 s81, v32, 25
; VI-NEXT:    v_readlane_b32 s80, v32, 24
; VI-NEXT:    v_readlane_b32 s71, v32, 23
; VI-NEXT:    v_readlane_b32 s70, v32, 22
; VI-NEXT:    v_readlane_b32 s69, v32, 21
; VI-NEXT:    v_readlane_b32 s68, v32, 20
; VI-NEXT:    v_readlane_b32 s67, v32, 19
; VI-NEXT:    v_readlane_b32 s66, v32, 18
; VI-NEXT:    v_readlane_b32 s65, v32, 17
; VI-NEXT:    v_readlane_b32 s64, v32, 16
; VI-NEXT:    v_readlane_b32 s55, v32, 15
; VI-NEXT:    v_readlane_b32 s54, v32, 14
; VI-NEXT:    v_readlane_b32 s53, v32, 13
; VI-NEXT:    v_readlane_b32 s52, v32, 12
; VI-NEXT:    v_readlane_b32 s51, v32, 11
; VI-NEXT:    v_readlane_b32 s50, v32, 10
; VI-NEXT:    v_readlane_b32 s49, v32, 9
; VI-NEXT:    v_readlane_b32 s48, v32, 8
; VI-NEXT:    v_readlane_b32 s39, v32, 7
; VI-NEXT:    v_readlane_b32 s38, v32, 6
; VI-NEXT:    v_readlane_b32 s37, v32, 5
; VI-NEXT:    v_readlane_b32 s36, v32, 4
; VI-NEXT:    v_readlane_b32 s35, v32, 3
; VI-NEXT:    v_readlane_b32 s34, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB57_4:
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr71
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    v_writelane_b32 v33, s70, 0
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr82
; VI-NEXT:    ; implicit-def: $sgpr53
; VI-NEXT:    ; implicit-def: $sgpr67
; VI-NEXT:    ; implicit-def: $sgpr83
; VI-NEXT:    ; implicit-def: $sgpr85
; VI-NEXT:    ; implicit-def: $sgpr47
; VI-NEXT:    ; implicit-def: $sgpr81
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr87
; VI-NEXT:    ; implicit-def: $sgpr80
; VI-NEXT:    ; implicit-def: $sgpr69
; VI-NEXT:    ; implicit-def: $sgpr84
; VI-NEXT:    ; implicit-def: $sgpr54
; VI-NEXT:    ; implicit-def: $sgpr68
; VI-NEXT:    ; implicit-def: $sgpr65
; VI-NEXT:    ; implicit-def: $sgpr55
; VI-NEXT:    ; implicit-def: $sgpr50
; VI-NEXT:    ; implicit-def: $sgpr64
; VI-NEXT:    ; implicit-def: $sgpr66
; VI-NEXT:    ; implicit-def: $sgpr52
; VI-NEXT:    ; implicit-def: $sgpr86
; VI-NEXT:    ; implicit-def: $sgpr51
; VI-NEXT:    ; implicit-def: $sgpr48
; VI-NEXT:    ; implicit-def: $sgpr38
; VI-NEXT:    ; implicit-def: $sgpr36
; VI-NEXT:    ; implicit-def: $sgpr34
; VI-NEXT:    ; implicit-def: $sgpr30
; VI-NEXT:    ; implicit-def: $sgpr90
; VI-NEXT:    ; implicit-def: $sgpr88
; VI-NEXT:    ; implicit-def: $sgpr78
; VI-NEXT:    ; implicit-def: $sgpr76
; VI-NEXT:    ; implicit-def: $sgpr74
; VI-NEXT:    ; implicit-def: $sgpr72
; VI-NEXT:    ; implicit-def: $sgpr62
; VI-NEXT:    ; implicit-def: $sgpr60
; VI-NEXT:    ; implicit-def: $sgpr58
; VI-NEXT:    v_writelane_b32 v33, s71, 1
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; kill: killed $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    s_branch .LBB57_2
;
; GFX9-LABEL: bitcast_v16i64_to_v128i8_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v29, s30, 0
; GFX9-NEXT:    v_writelane_b32 v29, s31, 1
; GFX9-NEXT:    v_writelane_b32 v29, s34, 2
; GFX9-NEXT:    v_writelane_b32 v29, s35, 3
; GFX9-NEXT:    v_writelane_b32 v29, s36, 4
; GFX9-NEXT:    v_writelane_b32 v29, s37, 5
; GFX9-NEXT:    v_writelane_b32 v29, s38, 6
; GFX9-NEXT:    v_writelane_b32 v29, s39, 7
; GFX9-NEXT:    v_writelane_b32 v29, s48, 8
; GFX9-NEXT:    v_writelane_b32 v29, s49, 9
; GFX9-NEXT:    v_writelane_b32 v29, s50, 10
; GFX9-NEXT:    v_writelane_b32 v29, s51, 11
; GFX9-NEXT:    v_writelane_b32 v29, s52, 12
; GFX9-NEXT:    v_writelane_b32 v29, s53, 13
; GFX9-NEXT:    v_writelane_b32 v29, s54, 14
; GFX9-NEXT:    v_writelane_b32 v29, s55, 15
; GFX9-NEXT:    v_writelane_b32 v29, s64, 16
; GFX9-NEXT:    v_writelane_b32 v29, s65, 17
; GFX9-NEXT:    v_writelane_b32 v29, s66, 18
; GFX9-NEXT:    v_writelane_b32 v29, s67, 19
; GFX9-NEXT:    v_writelane_b32 v29, s68, 20
; GFX9-NEXT:    v_writelane_b32 v29, s69, 21
; GFX9-NEXT:    v_writelane_b32 v29, s70, 22
; GFX9-NEXT:    v_writelane_b32 v29, s71, 23
; GFX9-NEXT:    v_writelane_b32 v29, s80, 24
; GFX9-NEXT:    v_writelane_b32 v29, s81, 25
; GFX9-NEXT:    v_writelane_b32 v29, s82, 26
; GFX9-NEXT:    v_writelane_b32 v29, s83, 27
; GFX9-NEXT:    v_writelane_b32 v29, s84, 28
; GFX9-NEXT:    v_writelane_b32 v29, s85, 29
; GFX9-NEXT:    v_writelane_b32 v29, s86, 30
; GFX9-NEXT:    v_writelane_b32 v29, s87, 31
; GFX9-NEXT:    v_writelane_b32 v29, s96, 32
; GFX9-NEXT:    v_writelane_b32 v29, s97, 33
; GFX9-NEXT:    v_readfirstlane_b32 s44, v19
; GFX9-NEXT:    v_writelane_b32 v29, s98, 34
; GFX9-NEXT:    v_readfirstlane_b32 s5, v18
; GFX9-NEXT:    v_readfirstlane_b32 s4, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s6, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s8, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s10, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s12, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s14, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s40, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s42, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    s_cmp_lg_u32 s44, 0
; GFX9-NEXT:    v_readfirstlane_b32 s44, v1
; GFX9-NEXT:    v_writelane_b32 v29, s99, 35
; GFX9-NEXT:    ; implicit-def: $vgpr30 : SGPR spill to VGPR lane
; GFX9-NEXT:    s_cbranch_scc0 .LBB57_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 0
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s6, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s6, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s8, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s8, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s11, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s11, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s11, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s10, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s10, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s13, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s13, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s13, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s12, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s12, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s15, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s15, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s15, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s14, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s14, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s41, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s41, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s41, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s40, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s40, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s43, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s43, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s43, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s42, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s42, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s45, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s45, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s45, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s44, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s44, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s26, 16
; GFX9-NEXT:    s_lshr_b32 s85, s29, 8
; GFX9-NEXT:    s_lshr_b32 s96, s28, 8
; GFX9-NEXT:    s_lshr_b32 s71, s27, 24
; GFX9-NEXT:    s_lshr_b32 s51, s27, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 49
; GFX9-NEXT:    s_lshr_b32 s67, s26, 8
; GFX9-NEXT:    s_lshr_b32 s82, s25, 24
; GFX9-NEXT:    s_lshr_b32 s84, s25, 16
; GFX9-NEXT:    s_lshr_b32 s65, s25, 8
; GFX9-NEXT:    s_lshr_b32 s86, s24, 16
; GFX9-NEXT:    s_lshr_b32 s55, s24, 8
; GFX9-NEXT:    s_lshr_b32 s98, s23, 24
; GFX9-NEXT:    s_lshr_b32 s87, s23, 16
; GFX9-NEXT:    s_lshr_b32 s66, s23, 8
; GFX9-NEXT:    s_lshr_b32 s97, s22, 16
; GFX9-NEXT:    s_lshr_b32 s64, s22, 8
; GFX9-NEXT:    s_lshr_b32 s49, s21, 24
; GFX9-NEXT:    s_lshr_b32 s99, s21, 16
; GFX9-NEXT:    s_lshr_b32 s80, s21, 8
; GFX9-NEXT:    s_lshr_b32 s39, s20, 16
; GFX9-NEXT:    s_lshr_b32 s70, s20, 8
; GFX9-NEXT:    s_lshr_b32 s48, s19, 24
; GFX9-NEXT:    s_lshr_b32 s38, s19, 16
; GFX9-NEXT:    s_lshr_b32 s68, s19, 8
; GFX9-NEXT:    s_lshr_b32 s54, s18, 16
; GFX9-NEXT:    s_lshr_b32 s69, s18, 8
; GFX9-NEXT:    s_lshr_b32 s52, s17, 24
; GFX9-NEXT:    s_lshr_b32 s50, s17, 16
; GFX9-NEXT:    s_lshr_b32 s83, s17, 8
; GFX9-NEXT:    s_lshr_b32 s53, s16, 16
; GFX9-NEXT:    s_lshr_b32 s81, s16, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[16:17], 24
; GFX9-NEXT:    s_cbranch_execnz .LBB57_3
; GFX9-NEXT:  .LBB57_2: ; %cmp.true
; GFX9-NEXT:    s_add_u32 s16, s16, 3
; GFX9-NEXT:    s_addc_u32 s17, s17, 0
; GFX9-NEXT:    s_add_u32 s18, s18, 3
; GFX9-NEXT:    s_addc_u32 s19, s19, 0
; GFX9-NEXT:    s_add_u32 s20, s20, 3
; GFX9-NEXT:    s_addc_u32 s21, s21, 0
; GFX9-NEXT:    s_add_u32 s22, s22, 3
; GFX9-NEXT:    s_addc_u32 s23, s23, 0
; GFX9-NEXT:    s_add_u32 s24, s24, 3
; GFX9-NEXT:    s_addc_u32 s25, s25, 0
; GFX9-NEXT:    s_add_u32 s26, s26, 3
; GFX9-NEXT:    s_addc_u32 s27, s27, 0
; GFX9-NEXT:    s_add_u32 s28, s28, 3
; GFX9-NEXT:    s_addc_u32 s29, s29, 0
; GFX9-NEXT:    s_add_u32 s44, s44, 3
; GFX9-NEXT:    s_addc_u32 s45, s45, 0
; GFX9-NEXT:    s_add_u32 s42, s42, 3
; GFX9-NEXT:    s_addc_u32 s43, s43, 0
; GFX9-NEXT:    s_add_u32 s40, s40, 3
; GFX9-NEXT:    s_addc_u32 s41, s41, 0
; GFX9-NEXT:    s_add_u32 s14, s14, 3
; GFX9-NEXT:    s_addc_u32 s15, s15, 0
; GFX9-NEXT:    s_add_u32 s12, s12, 3
; GFX9-NEXT:    s_addc_u32 s13, s13, 0
; GFX9-NEXT:    s_add_u32 s10, s10, 3
; GFX9-NEXT:    s_addc_u32 s11, s11, 0
; GFX9-NEXT:    s_add_u32 s8, s8, 3
; GFX9-NEXT:    s_addc_u32 s9, s9, 0
; GFX9-NEXT:    s_add_u32 s6, s6, 3
; GFX9-NEXT:    s_addc_u32 s7, s7, 0
; GFX9-NEXT:    s_add_u32 s4, s4, 3
; GFX9-NEXT:    s_addc_u32 s5, s5, 0
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 0
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s6, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s6, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s8, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s8, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s11, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s11, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s11, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s10, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s10, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s13, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s13, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s13, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s12, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s12, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s15, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s15, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s15, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s14, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s14, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s41, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s41, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s41, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s40, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s40, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s43, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s43, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s43, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s42, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s42, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s45, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s45, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s45, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s44, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s44, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v30, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    v_writelane_b32 v30, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s26, 16
; GFX9-NEXT:    s_lshr_b32 s85, s29, 8
; GFX9-NEXT:    s_lshr_b32 s96, s28, 8
; GFX9-NEXT:    s_lshr_b32 s71, s27, 24
; GFX9-NEXT:    s_lshr_b32 s51, s27, 8
; GFX9-NEXT:    v_writelane_b32 v30, s46, 49
; GFX9-NEXT:    s_lshr_b32 s67, s26, 8
; GFX9-NEXT:    s_lshr_b32 s82, s25, 24
; GFX9-NEXT:    s_lshr_b32 s84, s25, 16
; GFX9-NEXT:    s_lshr_b32 s65, s25, 8
; GFX9-NEXT:    s_lshr_b32 s86, s24, 16
; GFX9-NEXT:    s_lshr_b32 s55, s24, 8
; GFX9-NEXT:    s_lshr_b32 s98, s23, 24
; GFX9-NEXT:    s_lshr_b32 s87, s23, 16
; GFX9-NEXT:    s_lshr_b32 s66, s23, 8
; GFX9-NEXT:    s_lshr_b32 s97, s22, 16
; GFX9-NEXT:    s_lshr_b32 s64, s22, 8
; GFX9-NEXT:    s_lshr_b32 s49, s21, 24
; GFX9-NEXT:    s_lshr_b32 s99, s21, 16
; GFX9-NEXT:    s_lshr_b32 s80, s21, 8
; GFX9-NEXT:    s_lshr_b32 s39, s20, 16
; GFX9-NEXT:    s_lshr_b32 s70, s20, 8
; GFX9-NEXT:    s_lshr_b32 s48, s19, 24
; GFX9-NEXT:    s_lshr_b32 s38, s19, 16
; GFX9-NEXT:    s_lshr_b32 s68, s19, 8
; GFX9-NEXT:    s_lshr_b32 s54, s18, 16
; GFX9-NEXT:    s_lshr_b32 s69, s18, 8
; GFX9-NEXT:    s_lshr_b32 s52, s17, 24
; GFX9-NEXT:    s_lshr_b32 s50, s17, 16
; GFX9-NEXT:    s_lshr_b32 s83, s17, 8
; GFX9-NEXT:    s_lshr_b32 s53, s16, 16
; GFX9-NEXT:    s_lshr_b32 s81, s16, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[16:17], 24
; GFX9-NEXT:  .LBB57_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v6, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v2, s81
; GFX9-NEXT:    v_perm_b32 v6, s53, v6, v1
; GFX9-NEXT:    v_perm_b32 v3, s16, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v23, s52
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v4, s83
; GFX9-NEXT:    v_perm_b32 v23, s50, v23, v1
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v6
; GFX9-NEXT:    v_perm_b32 v2, s17, v4, v1
; GFX9-NEXT:    v_mov_b32_e32 v10, s34
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v23
; GFX9-NEXT:    v_mov_b32_e32 v4, s69
; GFX9-NEXT:    v_perm_b32 v10, s54, v10, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_perm_b32 v5, s18, v4, v1
; GFX9-NEXT:    v_mov_b32_e32 v24, s48
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v7, s68
; GFX9-NEXT:    v_perm_b32 v24, s38, v24, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v5, v2
; GFX9-NEXT:    v_perm_b32 v4, s19, v7, v1
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v24
; GFX9-NEXT:    v_mov_b32_e32 v7, s70
; GFX9-NEXT:    v_perm_b32 v14, s39, v14, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX9-NEXT:    v_perm_b32 v8, s20, v7, v1
; GFX9-NEXT:    v_mov_b32_e32 v25, s49
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v14
; GFX9-NEXT:    v_mov_b32_e32 v9, s80
; GFX9-NEXT:    v_perm_b32 v25, s99, v25, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v8, v2
; GFX9-NEXT:    v_perm_b32 v7, s21, v9, v1
; GFX9-NEXT:    v_mov_b32_e32 v15, s94
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v25
; GFX9-NEXT:    v_mov_b32_e32 v9, s64
; GFX9-NEXT:    v_perm_b32 v15, s97, v15, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v7, v2
; GFX9-NEXT:    v_perm_b32 v11, s22, v9, v1
; GFX9-NEXT:    v_mov_b32_e32 v26, s98
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v15
; GFX9-NEXT:    v_mov_b32_e32 v12, s66
; GFX9-NEXT:    v_perm_b32 v26, s87, v26, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v11, v2
; GFX9-NEXT:    v_perm_b32 v9, s23, v12, v1
; GFX9-NEXT:    v_mov_b32_e32 v16, s92
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v26
; GFX9-NEXT:    v_mov_b32_e32 v12, s55
; GFX9-NEXT:    v_perm_b32 v16, s86, v16, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v9, v2
; GFX9-NEXT:    v_perm_b32 v13, s24, v12, v1
; GFX9-NEXT:    v_mov_b32_e32 v27, s82
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v16
; GFX9-NEXT:    v_mov_b32_e32 v17, s65
; GFX9-NEXT:    v_perm_b32 v27, s84, v27, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v13, v2
; GFX9-NEXT:    v_perm_b32 v12, s25, v17, v1
; GFX9-NEXT:    v_mov_b32_e32 v17, s90
; GFX9-NEXT:    v_readlane_b32 s16, v30, 49
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
; GFX9-NEXT:    v_mov_b32_e32 v18, s67
; GFX9-NEXT:    v_perm_b32 v17, s16, v17, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v12, v2
; GFX9-NEXT:    v_perm_b32 v18, s26, v18, v1
; GFX9-NEXT:    v_mov_b32_e32 v28, s71
; GFX9-NEXT:    v_readlane_b32 s16, v30, 48
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
; GFX9-NEXT:    v_mov_b32_e32 v19, s51
; GFX9-NEXT:    v_perm_b32 v28, s16, v28, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v18, v2
; GFX9-NEXT:    v_perm_b32 v19, s27, v19, v1
; GFX9-NEXT:    v_mov_b32_e32 v20, s88
; GFX9-NEXT:    v_readlane_b32 s16, v30, 47
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
; GFX9-NEXT:    v_mov_b32_e32 v21, s96
; GFX9-NEXT:    v_perm_b32 v3, s16, v20, v1
; GFX9-NEXT:    v_or_b32_e32 v2, v19, v2
; GFX9-NEXT:    v_perm_b32 v21, s28, v21, v1
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v21, v2
; GFX9-NEXT:    v_readlane_b32 s16, v30, 45
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 46
; GFX9-NEXT:    v_mov_b32_e32 v22, s85
; GFX9-NEXT:    v_perm_b32 v2, s16, v2, v1
; GFX9-NEXT:    v_perm_b32 v22, s29, v22, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v2, v22, v2
; GFX9-NEXT:    v_readlane_b32 s16, v30, 44
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_mov_b32_e32 v3, s78
; GFX9-NEXT:    v_readlane_b32 s16, v30, 43
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s44, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 42
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 40
; GFX9-NEXT:    v_mov_b32_e32 v3, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 41
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s45, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 39
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_mov_b32_e32 v3, s76
; GFX9-NEXT:    v_readlane_b32 s16, v30, 38
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s42, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 37
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 35
; GFX9-NEXT:    v_mov_b32_e32 v3, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 36
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s43, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 34
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_mov_b32_e32 v3, s74
; GFX9-NEXT:    v_readlane_b32 s16, v30, 33
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s40, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 32
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 30
; GFX9-NEXT:    v_mov_b32_e32 v3, s16
; GFX9-NEXT:    v_readlane_b32 s16, v30, 31
; GFX9-NEXT:    v_perm_b32 v3, s16, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s41, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s16, v30, 29
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    v_mov_b32_e32 v2, s16
; GFX9-NEXT:    v_perm_b32 v2, s14, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s72
; GFX9-NEXT:    v_readlane_b32 s14, v30, 28
; GFX9-NEXT:    v_perm_b32 v3, s14, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s14, v30, 27
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    v_mov_b32_e32 v2, s14
; GFX9-NEXT:    v_readlane_b32 s14, v30, 25
; GFX9-NEXT:    v_mov_b32_e32 v3, s14
; GFX9-NEXT:    v_readlane_b32 s14, v30, 26
; GFX9-NEXT:    v_perm_b32 v3, s14, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s15, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s14, v30, 24
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    v_mov_b32_e32 v2, s14
; GFX9-NEXT:    v_perm_b32 v2, s12, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s62
; GFX9-NEXT:    v_readlane_b32 s12, v30, 23
; GFX9-NEXT:    v_perm_b32 v3, s12, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s12, v30, 22
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    v_mov_b32_e32 v2, s12
; GFX9-NEXT:    v_readlane_b32 s12, v30, 20
; GFX9-NEXT:    v_mov_b32_e32 v3, s12
; GFX9-NEXT:    v_readlane_b32 s12, v30, 21
; GFX9-NEXT:    v_perm_b32 v3, s12, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s13, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s12, v30, 19
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    v_mov_b32_e32 v2, s12
; GFX9-NEXT:    v_perm_b32 v2, s10, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s60
; GFX9-NEXT:    v_readlane_b32 s10, v30, 18
; GFX9-NEXT:    v_perm_b32 v3, s10, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s10, v30, 17
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    v_mov_b32_e32 v2, s10
; GFX9-NEXT:    v_readlane_b32 s10, v30, 15
; GFX9-NEXT:    v_mov_b32_e32 v3, s10
; GFX9-NEXT:    v_readlane_b32 s10, v30, 16
; GFX9-NEXT:    v_perm_b32 v3, s10, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s11, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s10, v30, 14
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    v_mov_b32_e32 v2, s10
; GFX9-NEXT:    v_perm_b32 v2, s8, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s58
; GFX9-NEXT:    v_readlane_b32 s8, v30, 13
; GFX9-NEXT:    v_perm_b32 v3, s8, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s8, v30, 12
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    v_mov_b32_e32 v2, s8
; GFX9-NEXT:    v_readlane_b32 s8, v30, 10
; GFX9-NEXT:    v_mov_b32_e32 v3, s8
; GFX9-NEXT:    v_readlane_b32 s8, v30, 11
; GFX9-NEXT:    v_perm_b32 v3, s8, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s9, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s8, v30, 9
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    v_mov_b32_e32 v2, s8
; GFX9-NEXT:    v_perm_b32 v2, s6, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s56
; GFX9-NEXT:    v_readlane_b32 s6, v30, 8
; GFX9-NEXT:    v_perm_b32 v3, s6, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s6, v30, 7
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    v_mov_b32_e32 v2, s6
; GFX9-NEXT:    v_readlane_b32 s6, v30, 5
; GFX9-NEXT:    v_mov_b32_e32 v3, s6
; GFX9-NEXT:    v_readlane_b32 s6, v30, 6
; GFX9-NEXT:    v_perm_b32 v3, s6, v3, v1
; GFX9-NEXT:    v_perm_b32 v2, s7, v2, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s6, v30, 4
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    v_mov_b32_e32 v2, s6
; GFX9-NEXT:    v_perm_b32 v2, s4, v2, v1
; GFX9-NEXT:    v_mov_b32_e32 v3, s46
; GFX9-NEXT:    v_readlane_b32 s4, v30, 3
; GFX9-NEXT:    v_perm_b32 v3, s4, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_readlane_b32 s4, v30, 2
; GFX9-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    v_mov_b32_e32 v2, s4
; GFX9-NEXT:    v_readlane_b32 s4, v30, 0
; GFX9-NEXT:    v_mov_b32_e32 v3, s4
; GFX9-NEXT:    v_readlane_b32 s4, v30, 1
; GFX9-NEXT:    v_perm_b32 v2, s5, v2, v1
; GFX9-NEXT:    v_perm_b32 v1, s4, v3, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    v_readlane_b32 s99, v29, 35
; GFX9-NEXT:    v_readlane_b32 s98, v29, 34
; GFX9-NEXT:    v_readlane_b32 s97, v29, 33
; GFX9-NEXT:    v_readlane_b32 s96, v29, 32
; GFX9-NEXT:    v_readlane_b32 s87, v29, 31
; GFX9-NEXT:    v_readlane_b32 s86, v29, 30
; GFX9-NEXT:    v_readlane_b32 s85, v29, 29
; GFX9-NEXT:    v_readlane_b32 s84, v29, 28
; GFX9-NEXT:    v_readlane_b32 s83, v29, 27
; GFX9-NEXT:    v_readlane_b32 s82, v29, 26
; GFX9-NEXT:    v_readlane_b32 s81, v29, 25
; GFX9-NEXT:    v_readlane_b32 s80, v29, 24
; GFX9-NEXT:    v_readlane_b32 s71, v29, 23
; GFX9-NEXT:    v_readlane_b32 s70, v29, 22
; GFX9-NEXT:    v_readlane_b32 s69, v29, 21
; GFX9-NEXT:    v_readlane_b32 s68, v29, 20
; GFX9-NEXT:    v_readlane_b32 s67, v29, 19
; GFX9-NEXT:    v_readlane_b32 s66, v29, 18
; GFX9-NEXT:    v_readlane_b32 s65, v29, 17
; GFX9-NEXT:    v_readlane_b32 s64, v29, 16
; GFX9-NEXT:    v_readlane_b32 s55, v29, 15
; GFX9-NEXT:    v_readlane_b32 s54, v29, 14
; GFX9-NEXT:    v_readlane_b32 s53, v29, 13
; GFX9-NEXT:    v_readlane_b32 s52, v29, 12
; GFX9-NEXT:    v_readlane_b32 s51, v29, 11
; GFX9-NEXT:    v_readlane_b32 s50, v29, 10
; GFX9-NEXT:    v_readlane_b32 s49, v29, 9
; GFX9-NEXT:    v_readlane_b32 s48, v29, 8
; GFX9-NEXT:    v_readlane_b32 s39, v29, 7
; GFX9-NEXT:    v_readlane_b32 s38, v29, 6
; GFX9-NEXT:    v_readlane_b32 s37, v29, 5
; GFX9-NEXT:    v_readlane_b32 s36, v29, 4
; GFX9-NEXT:    v_readlane_b32 s35, v29, 3
; GFX9-NEXT:    v_readlane_b32 s34, v29, 2
; GFX9-NEXT:    v_readlane_b32 s31, v29, 1
; GFX9-NEXT:    v_readlane_b32 s30, v29, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB57_4:
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr81
; GFX9-NEXT:    ; implicit-def: $sgpr53
; GFX9-NEXT:    ; implicit-def: $sgpr83
; GFX9-NEXT:    ; implicit-def: $sgpr50
; GFX9-NEXT:    ; implicit-def: $sgpr52
; GFX9-NEXT:    ; implicit-def: $sgpr69
; GFX9-NEXT:    ; implicit-def: $sgpr54
; GFX9-NEXT:    ; implicit-def: $sgpr68
; GFX9-NEXT:    ; implicit-def: $sgpr38
; GFX9-NEXT:    ; implicit-def: $sgpr48
; GFX9-NEXT:    ; implicit-def: $sgpr70
; GFX9-NEXT:    ; implicit-def: $sgpr39
; GFX9-NEXT:    ; implicit-def: $sgpr80
; GFX9-NEXT:    ; implicit-def: $sgpr99
; GFX9-NEXT:    ; implicit-def: $sgpr49
; GFX9-NEXT:    ; implicit-def: $sgpr64
; GFX9-NEXT:    ; implicit-def: $sgpr97
; GFX9-NEXT:    ; implicit-def: $sgpr66
; GFX9-NEXT:    ; implicit-def: $sgpr87
; GFX9-NEXT:    ; implicit-def: $sgpr98
; GFX9-NEXT:    ; implicit-def: $sgpr55
; GFX9-NEXT:    ; implicit-def: $sgpr86
; GFX9-NEXT:    ; implicit-def: $sgpr65
; GFX9-NEXT:    ; implicit-def: $sgpr84
; GFX9-NEXT:    ; implicit-def: $sgpr82
; GFX9-NEXT:    ; implicit-def: $sgpr67
; GFX9-NEXT:    ; implicit-def: $sgpr51
; GFX9-NEXT:    ; implicit-def: $sgpr71
; GFX9-NEXT:    ; implicit-def: $sgpr96
; GFX9-NEXT:    ; implicit-def: $sgpr85
; GFX9-NEXT:    ; implicit-def: $sgpr36
; GFX9-NEXT:    ; implicit-def: $sgpr34
; GFX9-NEXT:    ; implicit-def: $sgpr30
; GFX9-NEXT:    ; implicit-def: $sgpr94
; GFX9-NEXT:    ; implicit-def: $sgpr92
; GFX9-NEXT:    ; implicit-def: $sgpr90
; GFX9-NEXT:    ; implicit-def: $sgpr88
; GFX9-NEXT:    ; implicit-def: $sgpr78
; GFX9-NEXT:    ; implicit-def: $sgpr76
; GFX9-NEXT:    ; implicit-def: $sgpr74
; GFX9-NEXT:    ; implicit-def: $sgpr72
; GFX9-NEXT:    ; implicit-def: $sgpr62
; GFX9-NEXT:    ; implicit-def: $sgpr60
; GFX9-NEXT:    ; implicit-def: $sgpr58
; GFX9-NEXT:    ; implicit-def: $sgpr56
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    s_branch .LBB57_2
;
; GFX11-LABEL: bitcast_v16i64_to_v128i8_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v24, s32
; GFX11-NEXT:    scratch_store_b32 off, v25, s32 offset:4
; GFX11-NEXT:    scratch_store_b32 off, v26, s32 offset:8
; GFX11-NEXT:    scratch_store_b32 off, v27, s32 offset:12
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v24, s30, 0
; GFX11-NEXT:    v_writelane_b32 v25, s96, 0
; GFX11-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_writelane_b32 v24, s31, 1
; GFX11-NEXT:    v_writelane_b32 v25, s97, 1
; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
; GFX11-NEXT:    v_writelane_b32 v24, s34, 2
; GFX11-NEXT:    v_writelane_b32 v25, s98, 2
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_writelane_b32 v24, s35, 3
; GFX11-NEXT:    v_writelane_b32 v25, s99, 3
; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
; GFX11-NEXT:    v_writelane_b32 v24, s36, 4
; GFX11-NEXT:    v_writelane_b32 v25, s100, 4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_writelane_b32 v24, s37, 5
; GFX11-NEXT:    v_writelane_b32 v25, s101, 5
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-NEXT:    ; implicit-def: $vgpr27 : SGPR spill to VGPR lane
; GFX11-NEXT:    ; implicit-def: $vgpr26 : SGPR spill to VGPR lane
; GFX11-NEXT:    v_writelane_b32 v24, s38, 6
; GFX11-NEXT:    v_writelane_b32 v25, s102, 6
; GFX11-NEXT:    v_writelane_b32 v24, s39, 7
; GFX11-NEXT:    v_writelane_b32 v25, s103, 7
; GFX11-NEXT:    v_writelane_b32 v24, s48, 8
; GFX11-NEXT:    v_writelane_b32 v25, s104, 8
; GFX11-NEXT:    v_writelane_b32 v24, s49, 9
; GFX11-NEXT:    v_writelane_b32 v24, s50, 10
; GFX11-NEXT:    v_writelane_b32 v24, s51, 11
; GFX11-NEXT:    v_writelane_b32 v24, s52, 12
; GFX11-NEXT:    v_writelane_b32 v24, s53, 13
; GFX11-NEXT:    v_writelane_b32 v24, s54, 14
; GFX11-NEXT:    v_writelane_b32 v24, s55, 15
; GFX11-NEXT:    v_writelane_b32 v24, s64, 16
; GFX11-NEXT:    v_writelane_b32 v24, s65, 17
; GFX11-NEXT:    v_writelane_b32 v24, s66, 18
; GFX11-NEXT:    v_writelane_b32 v24, s67, 19
; GFX11-NEXT:    v_writelane_b32 v24, s68, 20
; GFX11-NEXT:    v_writelane_b32 v24, s69, 21
; GFX11-NEXT:    v_writelane_b32 v24, s70, 22
; GFX11-NEXT:    v_writelane_b32 v24, s71, 23
; GFX11-NEXT:    v_writelane_b32 v24, s80, 24
; GFX11-NEXT:    v_writelane_b32 v24, s81, 25
; GFX11-NEXT:    v_writelane_b32 v24, s82, 26
; GFX11-NEXT:    v_writelane_b32 v24, s83, 27
; GFX11-NEXT:    v_writelane_b32 v24, s84, 28
; GFX11-NEXT:    v_writelane_b32 v24, s85, 29
; GFX11-NEXT:    v_writelane_b32 v24, s86, 30
; GFX11-NEXT:    v_writelane_b32 v24, s87, 31
; GFX11-NEXT:    s_cbranch_scc0 .LBB57_4
; GFX11-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s27, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
; GFX11-NEXT:    s_lshr_b32 s34, s27, 16
; GFX11-NEXT:    s_lshr_b32 s35, s27, 8
; GFX11-NEXT:    s_lshr_b32 s37, s26, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
; GFX11-NEXT:    s_lshr_b32 s36, s26, 8
; GFX11-NEXT:    s_lshr_b32 s39, s25, 24
; GFX11-NEXT:    s_lshr_b32 s48, s25, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s4, 16
; GFX11-NEXT:    s_lshr_b32 s38, s25, 8
; GFX11-NEXT:    s_lshr_b32 s54, s24, 16
; GFX11-NEXT:    s_lshr_b32 s49, s24, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s4, 8
; GFX11-NEXT:    s_lshr_b32 s50, s23, 24
; GFX11-NEXT:    s_lshr_b32 s52, s23, 16
; GFX11-NEXT:    s_lshr_b32 s51, s23, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
; GFX11-NEXT:    s_lshr_b32 s80, s22, 16
; GFX11-NEXT:    s_lshr_b32 s53, s22, 8
; GFX11-NEXT:    s_lshr_b32 s55, s21, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
; GFX11-NEXT:    s_lshr_b32 s64, s21, 16
; GFX11-NEXT:    s_lshr_b32 s69, s21, 8
; GFX11-NEXT:    s_lshr_b32 s84, s20, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
; GFX11-NEXT:    s_lshr_b32 s82, s20, 8
; GFX11-NEXT:    s_lshr_b32 s65, s19, 24
; GFX11-NEXT:    s_lshr_b32 s66, s19, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s6, 16
; GFX11-NEXT:    s_lshr_b32 s85, s19, 8
; GFX11-NEXT:    s_lshr_b32 s86, s18, 16
; GFX11-NEXT:    s_lshr_b32 s87, s18, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s6, 8
; GFX11-NEXT:    s_lshr_b32 s67, s17, 24
; GFX11-NEXT:    s_lshr_b32 s68, s17, 16
; GFX11-NEXT:    s_lshr_b32 s96, s17, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
; GFX11-NEXT:    s_lshr_b32 s97, s16, 16
; GFX11-NEXT:    s_lshr_b32 s98, s16, 8
; GFX11-NEXT:    s_lshr_b32 s70, s3, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
; GFX11-NEXT:    s_lshr_b32 s71, s3, 16
; GFX11-NEXT:    s_lshr_b32 s99, s3, 8
; GFX11-NEXT:    s_lshr_b32 s100, s2, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
; GFX11-NEXT:    s_lshr_b32 s101, s2, 8
; GFX11-NEXT:    s_lshr_b32 s81, s1, 24
; GFX11-NEXT:    s_lshr_b32 s83, s1, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s8, 16
; GFX11-NEXT:    s_lshr_b32 s102, s1, 8
; GFX11-NEXT:    s_lshr_b32 s103, s0, 16
; GFX11-NEXT:    s_lshr_b32 s104, s0, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s8, 8
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[6:7], 24
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[10:11], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s11, 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[24:25], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s11, 8
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[18:19], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s10, 16
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[16:17], 24
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[2:3], 24
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s10, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s13, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s12, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s12, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s15, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s15, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s15, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s14, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s14, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s41, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s41, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s41, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s40, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s40, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 7
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[4:5], 24
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-NEXT:    s_cbranch_vccnz .LBB57_3
; GFX11-NEXT:  .LBB57_2: ; %cmp.true
; GFX11-NEXT:    s_add_u32 s0, s0, 3
; GFX11-NEXT:    s_addc_u32 s1, s1, 0
; GFX11-NEXT:    s_add_u32 s2, s2, 3
; GFX11-NEXT:    s_addc_u32 s3, s3, 0
; GFX11-NEXT:    s_add_u32 s16, s16, 3
; GFX11-NEXT:    s_addc_u32 s17, s17, 0
; GFX11-NEXT:    s_add_u32 s18, s18, 3
; GFX11-NEXT:    s_addc_u32 s19, s19, 0
; GFX11-NEXT:    s_add_u32 s20, s20, 3
; GFX11-NEXT:    s_addc_u32 s21, s21, 0
; GFX11-NEXT:    s_add_u32 s22, s22, 3
; GFX11-NEXT:    s_addc_u32 s23, s23, 0
; GFX11-NEXT:    s_add_u32 s24, s24, 3
; GFX11-NEXT:    s_addc_u32 s25, s25, 0
; GFX11-NEXT:    s_add_u32 s26, s26, 3
; GFX11-NEXT:    s_addc_u32 s27, s27, 0
; GFX11-NEXT:    s_add_u32 s28, s28, 3
; GFX11-NEXT:    s_addc_u32 s29, s29, 0
; GFX11-NEXT:    s_add_u32 s40, s40, 3
; GFX11-NEXT:    s_addc_u32 s41, s41, 0
; GFX11-NEXT:    s_add_u32 s14, s14, 3
; GFX11-NEXT:    s_addc_u32 s15, s15, 0
; GFX11-NEXT:    s_add_u32 s12, s12, 3
; GFX11-NEXT:    s_addc_u32 s13, s13, 0
; GFX11-NEXT:    s_add_u32 s10, s10, 3
; GFX11-NEXT:    s_addc_u32 s11, s11, 0
; GFX11-NEXT:    s_add_u32 s8, s8, 3
; GFX11-NEXT:    s_addc_u32 s9, s9, 0
; GFX11-NEXT:    s_add_u32 s6, s6, 3
; GFX11-NEXT:    s_addc_u32 s7, s7, 0
; GFX11-NEXT:    s_add_u32 s4, s4, 3
; GFX11-NEXT:    s_addc_u32 s5, s5, 0
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s27, 24
; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
; GFX11-NEXT:    s_lshr_b32 s34, s27, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
; GFX11-NEXT:    s_lshr_b32 s35, s27, 8
; GFX11-NEXT:    s_lshr_b32 s37, s26, 16
; GFX11-NEXT:    s_lshr_b32 s36, s26, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
; GFX11-NEXT:    s_lshr_b32 s39, s25, 24
; GFX11-NEXT:    s_lshr_b32 s48, s25, 16
; GFX11-NEXT:    s_lshr_b32 s38, s25, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s4, 16
; GFX11-NEXT:    s_lshr_b32 s54, s24, 16
; GFX11-NEXT:    s_lshr_b32 s49, s24, 8
; GFX11-NEXT:    s_lshr_b32 s50, s23, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s4, 8
; GFX11-NEXT:    s_lshr_b32 s52, s23, 16
; GFX11-NEXT:    s_lshr_b32 s51, s23, 8
; GFX11-NEXT:    s_lshr_b32 s80, s22, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
; GFX11-NEXT:    s_lshr_b32 s53, s22, 8
; GFX11-NEXT:    s_lshr_b32 s55, s21, 24
; GFX11-NEXT:    s_lshr_b32 s64, s21, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
; GFX11-NEXT:    s_lshr_b32 s69, s21, 8
; GFX11-NEXT:    s_lshr_b32 s84, s20, 16
; GFX11-NEXT:    s_lshr_b32 s82, s20, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
; GFX11-NEXT:    s_lshr_b32 s65, s19, 24
; GFX11-NEXT:    s_lshr_b32 s66, s19, 16
; GFX11-NEXT:    s_lshr_b32 s85, s19, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s6, 16
; GFX11-NEXT:    s_lshr_b32 s86, s18, 16
; GFX11-NEXT:    s_lshr_b32 s87, s18, 8
; GFX11-NEXT:    s_lshr_b32 s67, s17, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s6, 8
; GFX11-NEXT:    s_lshr_b32 s68, s17, 16
; GFX11-NEXT:    s_lshr_b32 s96, s17, 8
; GFX11-NEXT:    s_lshr_b32 s97, s16, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
; GFX11-NEXT:    s_lshr_b32 s98, s16, 8
; GFX11-NEXT:    s_lshr_b32 s70, s3, 24
; GFX11-NEXT:    s_lshr_b32 s71, s3, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
; GFX11-NEXT:    s_lshr_b32 s99, s3, 8
; GFX11-NEXT:    s_lshr_b32 s100, s2, 16
; GFX11-NEXT:    s_lshr_b32 s101, s2, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
; GFX11-NEXT:    s_lshr_b32 s81, s1, 24
; GFX11-NEXT:    s_lshr_b32 s83, s1, 16
; GFX11-NEXT:    s_lshr_b32 s102, s1, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s8, 16
; GFX11-NEXT:    s_lshr_b32 s103, s0, 16
; GFX11-NEXT:    s_lshr_b32 s104, s0, 8
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[6:7], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s8, 8
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[10:11], 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[12:13], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s11, 24
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[14:15], 24
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[28:29], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[26:27], 24
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[24:25], 24
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[22:23], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s11, 8
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[18:19], 24
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[16:17], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s10, 16
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[2:3], 24
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[0:1], 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s10, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s13, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s12, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s12, 8
; GFX11-NEXT:    v_writelane_b32 v27, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s15, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s15, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s15, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s14, 16
; GFX11-NEXT:    v_writelane_b32 v27, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s14, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s41, 24
; GFX11-NEXT:    v_writelane_b32 v27, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s41, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v27, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s41, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s40, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s40, 8
; GFX11-NEXT:    v_writelane_b32 v26, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s28, 16
; GFX11-NEXT:    v_writelane_b32 v26, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s28, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v26, s42, 7
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[4:5], 24
; GFX11-NEXT:  .LBB57_3: ; %end
; GFX11-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; GFX11-NEXT:    v_readlane_b32 s31, v24, 1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_perm_b32 v2, s103, s30, v1
; GFX11-NEXT:    v_readlane_b32 s103, v25, 7
; GFX11-NEXT:    v_readlane_b32 s30, v24, 0
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-NEXT:    v_perm_b32 v19, s83, s81, v1
; GFX11-NEXT:    v_perm_b32 v3, s0, s104, v1
; GFX11-NEXT:    v_perm_b32 v4, s1, s102, v1
; GFX11-NEXT:    v_perm_b32 v5, s100, s92, v1
; GFX11-NEXT:    v_perm_b32 v20, s71, s70, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-NEXT:    v_perm_b32 v6, s2, s101, v1
; GFX11-NEXT:    v_perm_b32 v7, s3, s99, v1
; GFX11-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-NEXT:    v_perm_b32 v8, s97, s90, v1
; GFX11-NEXT:    v_or_b32_e32 v3, v4, v19
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v20
; GFX11-NEXT:    v_perm_b32 v11, s86, s88, v1
; GFX11-NEXT:    v_perm_b32 v21, s68, s67, v1
; GFX11-NEXT:    v_perm_b32 v22, s66, s65, v1
; GFX11-NEXT:    v_or_b32_e32 v4, v6, v4
; GFX11-NEXT:    v_or_b32_e32 v5, v7, v5
; GFX11-NEXT:    v_perm_b32 v9, s16, s98, v1
; GFX11-NEXT:    v_perm_b32 v10, s17, s96, v1
; GFX11-NEXT:    v_perm_b32 v12, s18, s87, v1
; GFX11-NEXT:    v_perm_b32 v13, s19, s85, v1
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off
; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v8
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v21
; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 16, v11
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v22
; GFX11-NEXT:    v_perm_b32 v15, s80, s76, v1
; GFX11-NEXT:    v_perm_b32 v23, s64, s55, v1
; GFX11-NEXT:    v_or_b32_e32 v2, v9, v2
; GFX11-NEXT:    v_or_b32_e32 v3, v10, v3
; GFX11-NEXT:    v_or_b32_e32 v4, v12, v4
; GFX11-NEXT:    v_or_b32_e32 v5, v13, v5
; GFX11-NEXT:    v_perm_b32 v14, s84, s78, v1
; GFX11-NEXT:    v_perm_b32 v17, s21, s69, v1
; GFX11-NEXT:    v_perm_b32 v18, s54, s94, v1
; GFX11-NEXT:    v_perm_b32 v6, s22, s53, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v8, 16, v23
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v15
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:16
; GFX11-NEXT:    v_perm_b32 v5, s52, s50, v1
; GFX11-NEXT:    v_perm_b32 v16, s20, s82, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v14
; GFX11-NEXT:    v_or_b32_e32 v3, v17, v8
; GFX11-NEXT:    v_or_b32_e32 v4, v6, v9
; GFX11-NEXT:    v_perm_b32 v6, s23, s51, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-NEXT:    v_perm_b32 v8, s24, s49, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v18
; GFX11-NEXT:    v_perm_b32 v11, s37, s74, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 6
; GFX11-NEXT:    v_or_b32_e32 v2, v16, v7
; GFX11-NEXT:    v_perm_b32 v7, s48, s39, v1
; GFX11-NEXT:    v_or_b32_e32 v5, v6, v5
; GFX11-NEXT:    v_or_b32_e32 v6, v8, v9
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v11
; GFX11-NEXT:    v_perm_b32 v11, s0, s72, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 3
; GFX11-NEXT:    v_readlane_b32 s1, v26, 4
; GFX11-NEXT:    v_perm_b32 v10, s25, s38, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:32
; GFX11-NEXT:    v_perm_b32 v8, s26, s36, v1
; GFX11-NEXT:    v_perm_b32 v2, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 7
; GFX11-NEXT:    v_or_b32_e32 v7, v10, v7
; GFX11-NEXT:    v_perm_b32 v10, s34, vcc_hi, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v11
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-NEXT:    v_perm_b32 v4, s28, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 5
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v10
; GFX11-NEXT:    v_perm_b32 v9, s27, s35, v1
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s29, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 1
; GFX11-NEXT:    v_readlane_b32 s1, v27, 31
; GFX11-NEXT:    v_or_b32_e32 v9, v9, v3
; GFX11-NEXT:    v_readlane_b32 s104, v25, 8
; GFX11-NEXT:    v_or_b32_e32 v3, v10, v11
; GFX11-NEXT:    v_perm_b32 v12, s0, s62, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 2
; GFX11-NEXT:    scratch_store_b128 v0, v[6:9], off offset:48
; GFX11-NEXT:    v_readlane_b32 s102, v25, 6
; GFX11-NEXT:    v_readlane_b32 s101, v25, 5
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v12
; GFX11-NEXT:    v_perm_b32 v4, s40, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 30
; GFX11-NEXT:    v_readlane_b32 s100, v25, 4
; GFX11-NEXT:    v_readlane_b32 s99, v25, 3
; GFX11-NEXT:    v_readlane_b32 s98, v25, 2
; GFX11-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v26, 0
; GFX11-NEXT:    v_readlane_b32 s1, v27, 26
; GFX11-NEXT:    v_readlane_b32 s97, v25, 1
; GFX11-NEXT:    v_readlane_b32 s96, v25, 0
; GFX11-NEXT:    v_lshlrev_b32_e32 v7, 16, v10
; GFX11-NEXT:    v_perm_b32 v5, s41, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 28
; GFX11-NEXT:    v_readlane_b32 s87, v24, 31
; GFX11-NEXT:    v_readlane_b32 s86, v24, 30
; GFX11-NEXT:    v_readlane_b32 s85, v24, 29
; GFX11-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX11-NEXT:    v_perm_b32 v11, s0, s60, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 25
; GFX11-NEXT:    v_readlane_b32 s84, v24, 28
; GFX11-NEXT:    v_readlane_b32 s83, v24, 27
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:64
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v11
; GFX11-NEXT:    v_perm_b32 v6, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 29
; GFX11-NEXT:    v_readlane_b32 s1, v27, 21
; GFX11-NEXT:    v_readlane_b32 s82, v24, 26
; GFX11-NEXT:    v_readlane_b32 s81, v24, 25
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v6
; GFX11-NEXT:    v_perm_b32 v8, s14, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 27
; GFX11-NEXT:    v_readlane_b32 s80, v24, 24
; GFX11-NEXT:    v_readlane_b32 s71, v24, 23
; GFX11-NEXT:    v_readlane_b32 s70, v24, 22
; GFX11-NEXT:    v_or_b32_e32 v6, v8, v9
; GFX11-NEXT:    v_perm_b32 v10, s15, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 23
; GFX11-NEXT:    v_readlane_b32 s69, v24, 21
; GFX11-NEXT:    v_readlane_b32 s68, v24, 20
; GFX11-NEXT:    v_readlane_b32 s67, v24, 19
; GFX11-NEXT:    v_or_b32_e32 v7, v10, v11
; GFX11-NEXT:    v_perm_b32 v12, s0, s58, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 24
; GFX11-NEXT:    v_readlane_b32 s66, v24, 18
; GFX11-NEXT:    v_readlane_b32 s65, v24, 17
; GFX11-NEXT:    v_readlane_b32 s64, v24, 16
; GFX11-NEXT:    v_lshlrev_b32_e32 v9, 16, v12
; GFX11-NEXT:    v_perm_b32 v8, s12, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 20
; GFX11-NEXT:    v_readlane_b32 s55, v24, 15
; GFX11-NEXT:    v_readlane_b32 s54, v24, 14
; GFX11-NEXT:    v_readlane_b32 s53, v24, 13
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 22
; GFX11-NEXT:    v_readlane_b32 s1, v27, 16
; GFX11-NEXT:    v_readlane_b32 s52, v24, 12
; GFX11-NEXT:    v_readlane_b32 s51, v24, 11
; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v10
; GFX11-NEXT:    v_perm_b32 v9, s13, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 18
; GFX11-NEXT:    v_readlane_b32 s50, v24, 10
; GFX11-NEXT:    v_readlane_b32 s49, v24, 9
; GFX11-NEXT:    v_readlane_b32 s48, v24, 8
; GFX11-NEXT:    v_or_b32_e32 v9, v9, v3
; GFX11-NEXT:    v_perm_b32 v11, s0, s56, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 15
; GFX11-NEXT:    v_readlane_b32 s39, v24, 7
; GFX11-NEXT:    v_readlane_b32 s38, v24, 6
; GFX11-NEXT:    v_readlane_b32 s37, v24, 5
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v11
; GFX11-NEXT:    v_perm_b32 v2, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 19
; GFX11-NEXT:    v_readlane_b32 s1, v27, 11
; GFX11-NEXT:    v_readlane_b32 s36, v24, 4
; GFX11-NEXT:    v_readlane_b32 s35, v24, 3
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v2
; GFX11-NEXT:    v_perm_b32 v4, s10, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 17
; GFX11-NEXT:    v_readlane_b32 s34, v24, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s11, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 13
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v3, v10, v11
; GFX11-NEXT:    v_perm_b32 v12, s0, s46, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 14
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v12
; GFX11-NEXT:    v_perm_b32 v4, s8, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 10
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 12
; GFX11-NEXT:    v_readlane_b32 s1, v27, 6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_perm_b32 v5, s9, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 8
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v5, v5, v10
; GFX11-NEXT:    v_perm_b32 v11, s0, s44, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-NEXT:    v_perm_b32 v10, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 9
; GFX11-NEXT:    v_readlane_b32 s1, v27, 1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v10
; GFX11-NEXT:    v_perm_b32 v12, s6, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 7
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v10, v12, v11
; GFX11-NEXT:    v_perm_b32 v14, s7, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v11, v14, v13
; GFX11-NEXT:    v_perm_b32 v12, s0, s42, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-NEXT:    v_perm_b32 v13, s4, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX11-NEXT:    v_perm_b32 v13, s1, s0, v1
; GFX11-NEXT:    v_readlane_b32 s0, v27, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-NEXT:    v_perm_b32 v1, s5, s0, v1
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_or_b32_e32 v13, v1, v13
; GFX11-NEXT:    s_clause 0x2
; GFX11-NEXT:    scratch_store_b128 v0, v[6:9], off offset:80
; GFX11-NEXT:    scratch_store_b128 v0, v[2:5], off offset:96
; GFX11-NEXT:    scratch_store_b128 v0, v[10:13], off offset:112
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v24, off, s32
; GFX11-NEXT:    scratch_load_b32 v25, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v26, off, s32 offset:8
; GFX11-NEXT:    scratch_load_b32 v27, off, s32 offset:12
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB57_4:
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr104
; GFX11-NEXT:    ; implicit-def: $sgpr103
; GFX11-NEXT:    ; implicit-def: $sgpr30
; GFX11-NEXT:    ; implicit-def: $sgpr102
; GFX11-NEXT:    ; implicit-def: $sgpr83
; GFX11-NEXT:    ; implicit-def: $sgpr81
; GFX11-NEXT:    ; implicit-def: $sgpr101
; GFX11-NEXT:    ; implicit-def: $sgpr100
; GFX11-NEXT:    ; implicit-def: $sgpr92
; GFX11-NEXT:    ; implicit-def: $sgpr99
; GFX11-NEXT:    ; implicit-def: $sgpr71
; GFX11-NEXT:    ; implicit-def: $sgpr70
; GFX11-NEXT:    ; implicit-def: $sgpr98
; GFX11-NEXT:    ; implicit-def: $sgpr97
; GFX11-NEXT:    ; implicit-def: $sgpr90
; GFX11-NEXT:    ; implicit-def: $sgpr96
; GFX11-NEXT:    ; implicit-def: $sgpr68
; GFX11-NEXT:    ; implicit-def: $sgpr67
; GFX11-NEXT:    ; implicit-def: $sgpr87
; GFX11-NEXT:    ; implicit-def: $sgpr86
; GFX11-NEXT:    ; implicit-def: $sgpr88
; GFX11-NEXT:    ; implicit-def: $sgpr85
; GFX11-NEXT:    ; implicit-def: $sgpr66
; GFX11-NEXT:    ; implicit-def: $sgpr65
; GFX11-NEXT:    ; implicit-def: $sgpr82
; GFX11-NEXT:    ; implicit-def: $sgpr84
; GFX11-NEXT:    ; implicit-def: $sgpr78
; GFX11-NEXT:    ; implicit-def: $sgpr69
; GFX11-NEXT:    ; implicit-def: $sgpr64
; GFX11-NEXT:    ; implicit-def: $sgpr55
; GFX11-NEXT:    ; implicit-def: $sgpr53
; GFX11-NEXT:    ; implicit-def: $sgpr80
; GFX11-NEXT:    ; implicit-def: $sgpr76
; GFX11-NEXT:    ; implicit-def: $sgpr51
; GFX11-NEXT:    ; implicit-def: $sgpr52
; GFX11-NEXT:    ; implicit-def: $sgpr50
; GFX11-NEXT:    ; implicit-def: $sgpr49
; GFX11-NEXT:    ; implicit-def: $sgpr54
; GFX11-NEXT:    ; implicit-def: $sgpr38
; GFX11-NEXT:    ; implicit-def: $sgpr48
; GFX11-NEXT:    ; implicit-def: $sgpr39
; GFX11-NEXT:    ; implicit-def: $sgpr36
; GFX11-NEXT:    ; implicit-def: $sgpr37
; GFX11-NEXT:    ; implicit-def: $sgpr35
; GFX11-NEXT:    ; implicit-def: $sgpr34
; GFX11-NEXT:    ; implicit-def: $vcc_hi
; GFX11-NEXT:    ; implicit-def: $sgpr94
; GFX11-NEXT:    ; implicit-def: $sgpr74
; GFX11-NEXT:    ; implicit-def: $sgpr72
; GFX11-NEXT:    ; implicit-def: $sgpr62
; GFX11-NEXT:    ; implicit-def: $sgpr60
; GFX11-NEXT:    ; implicit-def: $sgpr58
; GFX11-NEXT:    ; implicit-def: $sgpr56
; GFX11-NEXT:    ; implicit-def: $sgpr46
; GFX11-NEXT:    ; implicit-def: $sgpr44
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    s_branch .LBB57_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) {
; SI-LABEL: bitcast_v128i8_to_v16i64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:188
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:204
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:212
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:220
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:268
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:292
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:316
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:332
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:352
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:376
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:348
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:384
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380
; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:364
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:372
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB58_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    v_or_b32_e32 v2, v2, v41
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_or_b32_e32 v3, v40, v3
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v2, v2, v53
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v48
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v62
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v38
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v56
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v46
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
; SI-NEXT:    v_or_b32_e32 v8, v8, v44
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v52
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v43
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:  .LBB58_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB58_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_movk_i32 s6, 0x300
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b32 s7, 0x3000000
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v54, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v40, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v41, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v53, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v48, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v62, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v38, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v56, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v46, v7
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v8, v44, v8
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v33
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v59
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v18
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v27
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
; SI-NEXT:    v_or_b32_e32 v31, v43, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v31
; SI-NEXT:  .LBB58_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v16i64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:12
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:380
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB58_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s6, 0xc0c0004
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v9, v33, v62, s6
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_perm_b32 v10, v63, v60, s6
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v11, v53, v58, s6
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_perm_b32 v12, v51, v56, s6
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v37, v36, s6
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v45, v44, s6
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v49, v48, s6
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v0, v1, v0, s6
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s6
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v41, s6
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_perm_b32 v8, v55, v34, s6
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_perm_b32 v9, v61, v32, s6
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_perm_b32 v10, v59, v54, s6
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_perm_b32 v11, v57, v52, s6
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v47, v38, s6
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_perm_b32 v13, v35, v46, s6
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_perm_b32 v14, v43, v50, s6
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_perm_b32 v15, v39, v42, s6
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v30, v30, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v32, v40, v32, s6
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:  .LBB58_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB58_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v31, 0x300
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v41
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v62
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v60
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v58
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v56
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v36
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v48
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    s_waitcnt vmcnt(11)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    s_waitcnt vmcnt(10)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    s_waitcnt vmcnt(9)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    s_waitcnt vmcnt(8)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    s_waitcnt vmcnt(7)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    s_waitcnt vmcnt(6)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_add_u16_e32 v8, 3, v55
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v34
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 3, v33
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_add_u16_e32 v9, 3, v61
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v32
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 3, v63
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
; VI-NEXT:    v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_add_u16_e32 v10, 3, v59
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v54
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 3, v53
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_add_u16_e32 v11, 3, v57
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v52
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 3, v51
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_add_u16_e32 v12, 3, v47
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 3, v37
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
; VI-NEXT:    v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_add_u16_e32 v13, 3, v35
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v46
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 3, v45
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_add_u16_e32 v14, 3, v43
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v50
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v15, 3, v49
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_add_u16_e32 v15, 3, v39
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v42
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v30, v32
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v33, 3, v40
; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:  .LBB58_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v16i64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:12
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:380
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB58_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0xc0c0004
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v10, v55, v32, s6
; GFX9-NEXT:    v_perm_b32 v11, v61, v60, s6
; GFX9-NEXT:    v_perm_b32 v12, v53, v38, s6
; GFX9-NEXT:    v_perm_b32 v13, v57, v56, s6
; GFX9-NEXT:    v_perm_b32 v14, v51, v34, s6
; GFX9-NEXT:    v_perm_b32 v15, v45, v50, s6
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v16, v43, v16, s6
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v9, s6
; GFX9-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX9-NEXT:    v_perm_b32 v8, v35, v42, s6
; GFX9-NEXT:    v_perm_b32 v9, v41, v48, s6
; GFX9-NEXT:    v_lshl_or_b32 v8, v9, 16, v8
; GFX9-NEXT:    v_perm_b32 v9, v39, v40, s6
; GFX9-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v10, v62, v63, s6
; GFX9-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX9-NEXT:    v_perm_b32 v11, v59, v54, s6
; GFX9-NEXT:    v_lshl_or_b32 v11, v12, 16, v11
; GFX9-NEXT:    v_perm_b32 v12, v58, v37, s6
; GFX9-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v47, v52, s6
; GFX9-NEXT:    v_lshl_or_b32 v13, v14, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v33, v46, s6
; GFX9-NEXT:    v_lshl_or_b32 v14, v15, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v49, v44, s6
; GFX9-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s6
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    v_lshl_or_b32 v16, v17, 16, v16
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    v_lshl_or_b32 v18, v19, 16, v18
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    v_lshl_or_b32 v19, v20, 16, v19
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    v_lshl_or_b32 v21, v22, 16, v21
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    v_lshl_or_b32 v23, v24, 16, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    v_lshl_or_b32 v24, v25, 16, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    v_lshl_or_b32 v25, v26, 16, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    v_lshl_or_b32 v26, v27, 16, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    v_lshl_or_b32 v28, v29, 16, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    v_lshl_or_b32 v29, v30, 16, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    v_lshl_or_b32 v30, v31, 16, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v32, v36, v32, s6
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    v_lshl_or_b32 v31, v32, 16, v31
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:  .LBB58_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB58_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    s_movk_i32 s6, 0x300
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v48
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v32
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v60
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v56
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v34
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v50
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(25)
; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(24)
; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
; GFX9-NEXT:    s_waitcnt vmcnt(22)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(19)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(14)
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(13)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    s_waitcnt vmcnt(11)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    s_waitcnt vmcnt(8)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v35
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v42
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v41
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v39
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v40
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v55
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v63
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v61
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v54
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v53
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v58
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v37
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v57
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v52
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v51
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v33
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v46
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v45
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v49
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v44
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v43
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v32, 3, v36
; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
; GFX9-NEXT:  .LBB58_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v16i64:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:384
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:380
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:376
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:372
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:368
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:364
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:360
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:356
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:352
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:348
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:344
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:340
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:336
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:332
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:328
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:260
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:132
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:388
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:8
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v74.l, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v75.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v76.l, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v77.l, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v78.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v88.l, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v79.l, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v89.l, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v90.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v92.l, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v91.l, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v93.l, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v94.l, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v95.l, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v104.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v106.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v105.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v107.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v108.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v110.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v109.l, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v111.l, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v120.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v121.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v122.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v123.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v124.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v125.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v126.l, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v127.l, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB58_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v127, v126, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v125, v124, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v123, v122, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v121, v120, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v107, v105, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v106, v104, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v95, v94, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v111, v109, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v110, v108, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v93, v91, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v92, v90, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v89, v79, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v88, v78, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v77, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v75, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v74, v72, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v63, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v62, v60, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v59, v58, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v57, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v56, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v45, v43, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v44, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v41, v40, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v183, v181, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v182, v180, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v179, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v178, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v164, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v161, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v160, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v149, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v147, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v146, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v135, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v134, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v131, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v129, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v128, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v117, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v116, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v113, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v103, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v102, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v99, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v98, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v87, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v84, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v81, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v80, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v69, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v66, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v55, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v54, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v48, v49, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v37, v39, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v38, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v36, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v48, 16, v31
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v35, 16, v37
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v32, 16, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr127_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr126_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr125_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr124_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr122_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr121_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr120_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr111_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr109_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr108_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr107_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr105_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr106_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr104_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr95_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr94_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr93_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr91_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr92_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr90_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr89_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr88_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT:  .LBB58_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB58_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v127.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v125.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v123.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v0.h, 8, v126.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v124.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v122.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v121.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v0.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v120.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v111.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v110.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v109.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v3.h, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v108.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, v107.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, v106.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v95.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v3.l, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v3.h, v2.h
; GFX11-TRUE16-NEXT:    v_and_b16 v3.l, 0xff, v4.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v105.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v4.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v104.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, v92.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v90.l
; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v3.h, v3.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, v89.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v94.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v93.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, v88.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v77.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v5.l, 8, v91.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, v74.l, 3
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v72.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, v63.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, v62.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v5.l, v4.h
; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v6.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v79.l
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v78.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v59.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, v56.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v76.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v75.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v10.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v46.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, v45.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.h, 8, v73.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, v44.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v41.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, v182.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v180.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v7.h, v7.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v9.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v61.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v9.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v60.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, v179.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, v178.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v58.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v57.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v167.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, v164.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v162.l
; GFX11-TRUE16-NEXT:    v_or_b16 v9.l, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v10.l, 8, v47.l
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, v161.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, v160.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v149.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v9.h, v10.l, v9.h
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v11.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v43.l
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v42.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, v146.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v144.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, v135.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.l, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v40.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v183.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, v134.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v131.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.h, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v12.l, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.h, 8, v181.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, v128.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v118.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, v117.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, v116.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.h, v12.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v14.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v177.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v14.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v176.l
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v113.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.l, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, v102.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.h, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v166.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v165.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v23.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v100.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, v99.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v15.l, 8, v163.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, v98.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v87.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, v84.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v82.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v15.l, v14.h
; GFX11-TRUE16-NEXT:    v_or_b16 v15.l, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v16.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v151.l
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v150.l
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, v81.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v15.h, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, v80.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v148.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v147.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v69.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, v66.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v28.h, 8, v64.l
; GFX11-TRUE16-NEXT:    v_or_b16 v16.h, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v17.l, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.h, 8, v145.l
; GFX11-TRUE16-NEXT:    v_and_b16 v28.l, 0xff, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, v55.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v29.h, 8, v53.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.l, 8, v51.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.l, v17.h, v17.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.h, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v19.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v133.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v19.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v132.l
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.h, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, v52.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v18.l, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v18.h, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v130.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v129.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v28.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v50.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.l, v29.h, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.l, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.h, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v20.l, 8, v119.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, v38.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, v36.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v33.h, v34.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.h, 8, v37.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.h, v20.l, v19.h
; GFX11-TRUE16-NEXT:    v_or_b16 v20.l, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v21.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v115.l
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v114.l
; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v31.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.l, 8, v35.l
; GFX11-TRUE16-NEXT:    v_or_b16 v20.h, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v21.l, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v112.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v103.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT:    v_and_b16 v33.h, 0xff, v33.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.l, 8, v32.l
; GFX11-TRUE16-NEXT:    v_or_b16 v21.h, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v22.l, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.h, 8, v101.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.l, v34.l, v31.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.h, v33.l, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v32.l, v33.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.l, v22.h, v22.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.h, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v24.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v97.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v24.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v96.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v23.l, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v23.h, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v86.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v85.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v25.l, 8, v83.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v3.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v4.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v4.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v5.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v25.l, v24.h
; GFX11-TRUE16-NEXT:    v_or_b16 v25.l, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v26.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v71.l
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v70.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v6.l
; GFX11-TRUE16-NEXT:    v_or_b16 v25.h, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v6.h
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v68.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v67.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v7.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v7.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v8.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.h, 8, v65.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v8.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v9.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v27.l, v27.h, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v54.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v11.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v11.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v12.l
; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v12.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v13.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v14.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.h, v30.l, v27.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v28.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, v49.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v29.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, v39.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v30.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v48.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v14.h
; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v15.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v15.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v16.l
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v32.h, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v16.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v17.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v17.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v18.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v19.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v19.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v20.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v21.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v21.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v22.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v22.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v23.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v24.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v24.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v25.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v26.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v26.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v30.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v31.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v31.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v32.l
; GFX11-TRUE16-NEXT:  .LBB58_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v16i64:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v86, v0
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:384
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:380
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:376
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:372
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:368
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:364
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:360
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:356
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:352
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:348
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:344
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:340
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:336
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:332
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:328
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:260
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:188
; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v45, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v46, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v47, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v57, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v56, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v59, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v61, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v62, off, s32 offset:132
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v63, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v73, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v72, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v74, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v75, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v79, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v89, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v88, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v90, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v91, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v93, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v94, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v95, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v105, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_u16 v104, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v106, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v107, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v109, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v108, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v110, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v111, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v120, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_u16 v121, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v122, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v0, off, s32 offset:388
; GFX11-FAKE16-NEXT:    scratch_load_u16 v123, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v125, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v124, off, s32 offset:8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_u16 v126, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v127, off, s32
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, v30 :: v_dual_mov_b32 v32, v29
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v28 :: v_dual_mov_b32 v35, v27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, v26 :: v_dual_mov_b32 v36, v25
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, v24 :: v_dual_mov_b32 v39, v23
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v49, v22 :: v_dual_mov_b32 v48, v21
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v16 :: v_dual_mov_b32 v55, v15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v65, v14 :: v_dual_mov_b32 v64, v13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, v12 :: v_dual_mov_b32 v67, v11
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, v10 :: v_dual_mov_b32 v68, v9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, v8 :: v_dual_mov_b32 v71, v7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, v4 :: v_dual_mov_b32 v83, v3
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, v2 :: v_dual_mov_b32 v84, v1
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB58_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v86, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v85, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v80, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v66, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v65, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v54, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v69, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v52, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v50, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v49, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v38, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v37, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v33, v127, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v126, v124, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v125, v123, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v122, v121, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v120, v111, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v110, v108, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v109, v107, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v106, v104, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v105, v95, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v94, v92, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v93, v91, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v90, v88, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v89, v79, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v78, v77, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v76, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v74, v72, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v73, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v62, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v61, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v58, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v57, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v46, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v45, v43, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v42, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v40, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v182, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v181, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v178, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v177, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v166, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v165, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v162, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v161, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v150, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v148, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v146, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v145, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v134, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v133, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v130, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v129, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v118, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v117, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v115, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v112, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v103, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v34, v101, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v99, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v97, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v31
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v34, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v36, 16, v35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr127
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr126
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr124
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr125
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr122
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr121
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr120
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr111
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr108
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr109
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr107
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr106
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr105
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr95
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr93
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr91
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr89
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:  .LBB58_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB58_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v86, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v85, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v84
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v82, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v80, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v81
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v69, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v70, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v67
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v68
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v66, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v54, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v65, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v55
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v7, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v52, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v51
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v50, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v49, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v8, 8, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, v38, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, v37, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v34, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v33, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v10, 8, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v127
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v126, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v125, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v11, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v124
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v123
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v122, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v121
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v120, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v111
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v110, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v109, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v107
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v13, 8, v108
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, v106, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, v105, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v14, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v104
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v95
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v94, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v12, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v93, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v15, 8, v92
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v91
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v90, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v89, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v88
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v79
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v78, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v17, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v77
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v76, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v14, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v75
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v74, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v73, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v63
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v18, 8, v72
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, v62, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, v61, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v19, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v60
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v59
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v58, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v57, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v20, 8, v56
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v47
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v46, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v45, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v21, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v44
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v43
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v42, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v22, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v41
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v40, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v183
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v182, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v181, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v179
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v23, 8, v180
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, v178, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, v177, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v24, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v176
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v167
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v166, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v22, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v165, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v25, 8, v164
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v163
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v162, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v161, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v26, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v160
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v151
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v150, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v27, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v149
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v148, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v146, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v145, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v135
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v28, 8, v144
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v134, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v133, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v29, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v132
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v131
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, v130, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v129, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v30, 8, v128
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v119
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v26, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v118, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v117, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v31, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v115, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v116
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v33, 8, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v113
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v32, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, v112, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, v103, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, v101, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, v99, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v38, v97, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v34, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v102
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v37, 8, v100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v39, 8, v98
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v48, 8, v96
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v49, 8, v87
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v34, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v37, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v39, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v48, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v49, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v35, v36
; GFX11-FAKE16-NEXT:  .LBB58_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define inreg <16 x i64> @bitcast_v128i8_to_v16i64_scalar(<128 x i8> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v128i8_to_v16i64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s76, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s46, v30
; SI-NEXT:    v_readfirstlane_b32 s6, v29
; SI-NEXT:    v_readfirstlane_b32 s7, v28
; SI-NEXT:    v_readfirstlane_b32 s8, v27
; SI-NEXT:    v_readfirstlane_b32 s9, v26
; SI-NEXT:    v_readfirstlane_b32 s10, v25
; SI-NEXT:    v_readfirstlane_b32 s11, v24
; SI-NEXT:    v_readfirstlane_b32 s12, v23
; SI-NEXT:    v_readfirstlane_b32 s13, v22
; SI-NEXT:    v_readfirstlane_b32 s14, v21
; SI-NEXT:    v_readfirstlane_b32 s15, v20
; SI-NEXT:    v_readfirstlane_b32 s40, v19
; SI-NEXT:    v_readfirstlane_b32 s41, v18
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:96
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:104
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:328
; SI-NEXT:    v_readfirstlane_b32 s75, v1
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:332
; SI-NEXT:    v_readfirstlane_b32 s43, v16
; SI-NEXT:    v_readfirstlane_b32 s45, v14
; SI-NEXT:    v_readfirstlane_b32 s73, v3
; SI-NEXT:    v_readfirstlane_b32 s74, v2
; SI-NEXT:    v_readfirstlane_b32 s57, v11
; SI-NEXT:    v_readfirstlane_b32 s72, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v15
; SI-NEXT:    v_readfirstlane_b32 s47, v13
; SI-NEXT:    v_readfirstlane_b32 s56, v12
; SI-NEXT:    v_readfirstlane_b32 s58, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s60, v8
; SI-NEXT:    v_readfirstlane_b32 s61, v7
; SI-NEXT:    v_readfirstlane_b32 s62, v6
; SI-NEXT:    v_readfirstlane_b32 s63, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v50
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v51
; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v52
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v13, 24, v54
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v55
; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v40
; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v41
; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v42
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v43
; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v46
; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v56
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v59
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v61
; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v44
; SI-NEXT:    v_lshlrev_b32_e32 v52, 8, v45
; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v47
; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:316
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v30
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 24, v29
; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v28
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v26
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v25
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v24
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v23
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v22
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v20
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v19
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v18
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:268
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:260
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:252
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:292
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v34
; SI-NEXT:    v_lshlrev_b32_e32 v34, 8, v32
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:220
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204
; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 24, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v33
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:196
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:116
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB59_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v31
; SI-NEXT:    v_or_b32_e32 v0, s4, v0
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v12
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v11, v3
; SI-NEXT:    v_or_b32_e32 v11, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v30
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v13, v1
; SI-NEXT:    v_or_b32_e32 v13, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v9
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v14, v1
; SI-NEXT:    v_or_b32_e32 v14, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v42
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v0, v0, v60
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v15, v1
; SI-NEXT:    v_or_b32_e32 v15, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v63
; SI-NEXT:    v_or_b32_e32 v0, v0, v41
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v16, v1
; SI-NEXT:    v_or_b32_e32 v16, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v61
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v17
; SI-NEXT:    v_or_b32_e32 v0, v0, v40
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v55, v1
; SI-NEXT:    v_or_b32_e32 v17, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v44
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v53, v1
; SI-NEXT:    v_or_b32_e32 v18, v0, v1
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v51, v1
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v12, v2, v3
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s18, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s77, s19, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s77, s5
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s77, s21, 8
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s22, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s77, s78, s77
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s24, 0xff
; SI-NEXT:    s_lshl_b32 s78, s25, 8
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s26, 0xff
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s79, s27, 24
; SI-NEXT:    s_and_b32 s77, s77, 0xffff
; SI-NEXT:    s_or_b32 s78, s79, s78
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s28, 0xff
; SI-NEXT:    s_lshl_b32 s79, s29, 8
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s76, 0xff
; SI-NEXT:    s_lshl_b32 s79, s79, 16
; SI-NEXT:    s_lshl_b32 s88, s75, 24
; SI-NEXT:    s_and_b32 s78, s78, 0xffff
; SI-NEXT:    s_or_b32 s79, s88, s79
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s74, 0xff
; SI-NEXT:    s_lshl_b32 s88, s73, 8
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s72, 0xff
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s89, s63, 24
; SI-NEXT:    s_and_b32 s79, s79, 0xffff
; SI-NEXT:    s_or_b32 s88, s89, s88
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s62, 0xff
; SI-NEXT:    s_lshl_b32 s89, s61, 8
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s60, 0xff
; SI-NEXT:    s_lshl_b32 s89, s89, 16
; SI-NEXT:    s_lshl_b32 s90, s59, 24
; SI-NEXT:    s_and_b32 s88, s88, 0xffff
; SI-NEXT:    s_or_b32 s89, s90, s89
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s58, 0xff
; SI-NEXT:    s_lshl_b32 s90, s57, 8
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s56, 0xff
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_lshl_b32 s91, s47, 24
; SI-NEXT:    s_and_b32 s89, s89, 0xffff
; SI-NEXT:    s_or_b32 s90, s91, s90
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s45, 0xff
; SI-NEXT:    s_lshl_b32 s91, s44, 8
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s43, 0xff
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v52
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v19, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v49, v1
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    s_lshl_b32 s91, s91, 16
; SI-NEXT:    s_lshl_b32 s92, s42, 24
; SI-NEXT:    s_and_b32 s90, s90, 0xffff
; SI-NEXT:    s_or_b32 s91, s92, s91
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s41, 0xff
; SI-NEXT:    s_lshl_b32 s92, s40, 8
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s15, 0xff
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s93, s14, 24
; SI-NEXT:    s_and_b32 s91, s91, 0xffff
; SI-NEXT:    s_or_b32 s92, s93, s92
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s13, 0xff
; SI-NEXT:    s_lshl_b32 s93, s12, 8
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s11, 0xff
; SI-NEXT:    s_lshl_b32 s93, s93, 16
; SI-NEXT:    s_lshl_b32 s94, s10, 24
; SI-NEXT:    s_and_b32 s92, s92, 0xffff
; SI-NEXT:    s_or_b32 s93, s94, s93
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s9, 0xff
; SI-NEXT:    s_lshl_b32 s94, s8, 8
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    s_and_b32 s94, s7, 0xff
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_lshl_b32 s95, s6, 24
; SI-NEXT:    s_and_b32 s93, s93, 0xffff
; SI-NEXT:    s_or_b32 s94, s95, s94
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v3, s78
; SI-NEXT:    v_mov_b32_e32 v4, s79
; SI-NEXT:    v_mov_b32_e32 v5, s88
; SI-NEXT:    v_mov_b32_e32 v6, s89
; SI-NEXT:    v_mov_b32_e32 v7, s90
; SI-NEXT:    v_mov_b32_e32 v8, s91
; SI-NEXT:    v_mov_b32_e32 v9, s92
; SI-NEXT:    v_mov_b32_e32 v10, s93
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v52
; SI-NEXT:    v_or_b32_e32 v0, v0, v50
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v20, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v39, v1
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v48
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v21, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v37, v1
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v48
; SI-NEXT:    v_or_b32_e32 v0, v0, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v22, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v35, v1
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v38
; SI-NEXT:    v_or_b32_e32 v0, v0, v36
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v23, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v33, v1
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v36
; SI-NEXT:    v_or_b32_e32 v0, v0, v34
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v24, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    v_or_b32_e32 v0, v0, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v29, v1
; SI-NEXT:    v_or_b32_e32 v25, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v28, v1
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v0, v0, v59
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v26, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v58
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v56
; SI-NEXT:    v_or_b32_e32 v0, v0, v57
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v27, v1
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    v_or_b32_e32 v27, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v0, v0, v45
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_or_b32_e32 v28, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v46
; SI-NEXT:    v_mov_b32_e32 v45, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v60, v32
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v29, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v30, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v31, v0, v1
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s77
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_branch .LBB59_3
; SI-NEXT:  .LBB59_2:
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB59_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v43, v58
; SI-NEXT:    v_mov_b32_e32 v58, v33
; SI-NEXT:    v_mov_b32_e32 v33, v37
; SI-NEXT:    v_mov_b32_e32 v34, v39
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_cbranch_vccnz .LBB59_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_mov_b32_e32 v32, v35
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_and_b32 s17, s22, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s76, s76, 3
; SI-NEXT:    s_add_i32 s74, s74, 3
; SI-NEXT:    s_add_i32 s72, s72, 3
; SI-NEXT:    s_add_i32 s62, s62, 3
; SI-NEXT:    s_add_i32 s60, s60, 3
; SI-NEXT:    s_add_i32 s58, s58, 3
; SI-NEXT:    s_add_i32 s56, s56, 3
; SI-NEXT:    s_and_b32 s22, s56, 0xff
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_addk_i32 s12, 0x300
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    s_add_i32 s10, s10, 0x3000000
; SI-NEXT:    s_add_i32 s6, s6, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v9, s10
; SI-NEXT:    v_mov_b32_e32 v10, s6
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, s4, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, 0x3000000, v0
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s16, s18, 0xff
; SI-NEXT:    s_addk_i32 s4, 0x300
; SI-NEXT:    s_lshl_b32 s5, s19, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s5, s16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s16, s21, 8
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_addk_i32 s5, 0x300
; SI-NEXT:    s_lshl_b32 s16, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_and_b32 s16, s24, 0xff
; SI-NEXT:    s_lshl_b32 s17, s25, 8
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s18, s26, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s27, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s28, 0xff
; SI-NEXT:    s_lshl_b32 s18, s29, 8
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s19, s76, 0xff
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_lshl_b32 s18, s75, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s18, s74, 0xff
; SI-NEXT:    s_lshl_b32 s19, s73, 8
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s20, s72, 0xff
; SI-NEXT:    s_addk_i32 s18, 0x300
; SI-NEXT:    s_lshl_b32 s19, s63, 24
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s19, s19, s20
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s19, s62, 0xff
; SI-NEXT:    s_lshl_b32 s20, s61, 8
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s21, s60, 0xff
; SI-NEXT:    s_addk_i32 s19, 0x300
; SI-NEXT:    s_lshl_b32 s20, s59, 24
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s21
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s20, s58, 0xff
; SI-NEXT:    s_lshl_b32 s21, s57, 8
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_addk_i32 s20, 0x300
; SI-NEXT:    s_lshl_b32 s21, s47, 24
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_and_b32 s21, s45, 0xff
; SI-NEXT:    s_lshl_b32 s22, s44, 8
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    s_and_b32 s23, s43, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s22, s42, 24
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s23
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s22, s41, 0xff
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_lshl_b32 s23, s40, 8
; SI-NEXT:    s_or_b32 s22, s23, s22
; SI-NEXT:    s_addk_i32 s22, 0x300
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s14, s14, s22
; SI-NEXT:    s_add_i32 s4, s4, 0x3000000
; SI-NEXT:    s_add_i32 s5, s5, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 0x3000000
; SI-NEXT:    s_add_i32 s17, s17, 0x3000000
; SI-NEXT:    s_add_i32 s18, s18, 0x3000000
; SI-NEXT:    s_add_i32 s19, s19, 0x3000000
; SI-NEXT:    s_add_i32 s20, s20, 0x3000000
; SI-NEXT:    s_add_i32 s21, s21, 0x3000000
; SI-NEXT:    s_add_i32 s14, s14, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v4, s18
; SI-NEXT:    v_mov_b32_e32 v5, s19
; SI-NEXT:    v_mov_b32_e32 v6, s20
; SI-NEXT:    v_mov_b32_e32 v7, s21
; SI-NEXT:    v_mov_b32_e32 v8, s14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_add_i32_e32 v12, vcc, 0x3000000, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v39, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_mov_b32_e32 v3, s17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v13, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v37, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v42, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v41, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v44
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v55, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v54
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v53, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v51, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v48
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v34, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v38
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v33, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v32, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v58
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v62, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v60
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v59, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v26, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v57, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v27, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v45, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v28, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v0
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:  .LBB59_5: ; %end
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v16i64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s45, v15
; VI-NEXT:    v_readfirstlane_b32 s46, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s76, v0
; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_readfirstlane_b32 s6, v30
; VI-NEXT:    v_readfirstlane_b32 s7, v29
; VI-NEXT:    v_readfirstlane_b32 s8, v28
; VI-NEXT:    v_readfirstlane_b32 s9, v27
; VI-NEXT:    v_readfirstlane_b32 s10, v26
; VI-NEXT:    v_readfirstlane_b32 s11, v25
; VI-NEXT:    v_readfirstlane_b32 s12, v24
; VI-NEXT:    v_readfirstlane_b32 s13, v23
; VI-NEXT:    v_readfirstlane_b32 s14, v22
; VI-NEXT:    v_readfirstlane_b32 s15, v21
; VI-NEXT:    v_readfirstlane_b32 s40, v20
; VI-NEXT:    v_readfirstlane_b32 s41, v19
; VI-NEXT:    v_readfirstlane_b32 s42, v18
; VI-NEXT:    v_readfirstlane_b32 s43, v17
; VI-NEXT:    v_readfirstlane_b32 s44, v16
; VI-NEXT:    v_readfirstlane_b32 s56, v12
; VI-NEXT:    v_readfirstlane_b32 s57, v11
; VI-NEXT:    v_readfirstlane_b32 s58, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s60, v8
; VI-NEXT:    v_readfirstlane_b32 s61, v7
; VI-NEXT:    v_readfirstlane_b32 s62, v6
; VI-NEXT:    v_readfirstlane_b32 s63, v5
; VI-NEXT:    v_readfirstlane_b32 s72, v4
; VI-NEXT:    v_readfirstlane_b32 s73, v3
; VI-NEXT:    v_readfirstlane_b32 s74, v2
; VI-NEXT:    v_readfirstlane_b32 s75, v1
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; VI-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; VI-NEXT:    s_and_b64 s[4:5], vcc, exec
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    s_cbranch_scc0 .LBB59_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_mov_b32_e32 v3, s23
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_mov_b32_e32 v2, s21
; VI-NEXT:    v_perm_b32 v3, s22, v3, v11
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_perm_b32 v2, s20, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_or_b32_e32 v10, v10, v12
; VI-NEXT:    v_perm_b32 v12, v14, v13, s4
; VI-NEXT:    v_perm_b32 v11, s6, v15, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v52, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v54, v53, s4
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    v_perm_b32 v13, v48, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v50, v49, s4
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_perm_b32 v14, v36, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v38, v37, s4
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_perm_b32 v15, v32, v16, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v16, v34, v33, s4
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    v_perm_b32 v16, v60, v17, s4
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_perm_b32 v17, v62, v61, s4
; VI-NEXT:    v_or_b32_e32 v16, v17, v16
; VI-NEXT:    v_perm_b32 v17, v59, v18, s4
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_perm_b32 v18, v41, v40, s4
; VI-NEXT:    v_or_b32_e32 v17, v18, v17
; VI-NEXT:    v_perm_b32 v18, v57, v19, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v19, v42, v58, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_perm_b32 v19, v45, v20, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_perm_b32 v20, v56, v47, s4
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_perm_b32 v20, v44, v21, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_perm_b32 v21, v46, v26, s4
; VI-NEXT:    v_or_b32_e32 v20, v21, v20
; VI-NEXT:    v_perm_b32 v21, v31, v22, s4
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_perm_b32 v22, v24, v43, s4
; VI-NEXT:    v_or_b32_e32 v21, v22, v21
; VI-NEXT:    v_perm_b32 v22, v28, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_perm_b32 v23, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v22, v23, v22
; VI-NEXT:    v_perm_b32 v23, v25, v63, s4
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_perm_b32 v24, v27, v55, s4
; VI-NEXT:    v_or_b32_e32 v23, v24, v23
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    v_or_b32_e32 v24, v25, v24
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    v_or_b32_e32 v25, v26, v25
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    v_or_b32_e32 v26, v27, v26
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    v_or_b32_e32 v27, v28, v27
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v28, v29, v28
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    v_or_b32_e32 v29, v30, v29
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    v_or_b32_e32 v30, v31, v30
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v43, v46, v43, s4
; VI-NEXT:    v_or_b32_e32 v31, v43, v31
; VI-NEXT:    s_mov_b64 s[4:5], 0
; VI-NEXT:    s_branch .LBB59_3
; VI-NEXT:  .LBB59_2:
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    s_mov_b64 s[4:5], -1
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:  .LBB59_3: ; %Flow
; VI-NEXT:    v_mov_b32_e32 v43, v44
; VI-NEXT:    v_mov_b32_e32 v44, v45
; VI-NEXT:    v_mov_b32_e32 v45, v47
; VI-NEXT:    v_mov_b32_e32 v46, v57
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; VI-NEXT:    s_cbranch_vccnz .LBB59_5
; VI-NEXT:  ; %bb.4: ; %cmp.true
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    v_mov_b32_e32 v1, s21
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    v_mov_b32_e32 v2, s23
; VI-NEXT:    v_perm_b32 v1, s20, v1, v11
; VI-NEXT:    v_perm_b32 v2, s22, v2, v11
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x300, v2
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    s_add_i32 s76, s76, 3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x300, v3
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s74, s74, 3
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    s_add_i32 s72, s72, 3
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x300, v4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s62, s62, 3
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    s_add_i32 s60, s60, 3
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x300, v5
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s58, s58, 3
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    s_add_i32 s56, s56, 3
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x300, v6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x300, v7
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x300, v8
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x300, v9
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x300, v10
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v10, v12, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v48
; VI-NEXT:    v_perm_b32 v14, v14, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v36
; VI-NEXT:    v_perm_b32 v15, v15, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x3000000, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x3000000, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x3000000, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v6
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v7
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x3000000, v8
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x3000000, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v10
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v11, s6, v12, v11
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x300, v11
; VI-NEXT:    v_perm_b32 v16, v16, v17, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v60
; VI-NEXT:    v_perm_b32 v17, v17, v18, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v57
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v18, v18, v19, s4
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v12, v12, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v54
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v52
; VI-NEXT:    v_perm_b32 v12, v12, v53, s4
; VI-NEXT:    v_perm_b32 v13, v13, v51, s4
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x300, v12
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v50
; VI-NEXT:    v_perm_b32 v13, v13, v49, s4
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x300, v13
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v38
; VI-NEXT:    v_perm_b32 v14, v14, v37, s4
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x300, v14
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v34
; VI-NEXT:    v_perm_b32 v15, v15, v33, s4
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x300, v15
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v62
; VI-NEXT:    v_perm_b32 v16, v16, v61, s4
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x300, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v41
; VI-NEXT:    v_perm_b32 v17, v17, v40, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x300, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v42
; VI-NEXT:    v_perm_b32 v18, v18, v47, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v18
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v46
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x3000000, v11
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x3000000, v12
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x3000000, v13
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x3000000, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x3000000, v15
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x3000000, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x3000000, v17
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x3000000, v18
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_perm_b32 v20, v20, v21, s4
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v45
; VI-NEXT:    v_perm_b32 v20, v20, v44, s4
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x300, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x3000000, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v43
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x3000000, v21
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x3000000, v22
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x3000000, v23
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x3000000, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x3000000, v25
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x3000000, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x3000000, v27
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x3000000, v28
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x3000000, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x3000000, v30
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x300, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v32
; VI-NEXT:    v_perm_b32 v32, v32, v33, s4
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x3000000, v31
; VI-NEXT:  .LBB59_5: ; %end
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v16i64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s45, v15
; GFX9-NEXT:    v_readfirstlane_b32 s46, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s76, v0
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readfirstlane_b32 s6, v30
; GFX9-NEXT:    v_readfirstlane_b32 s7, v29
; GFX9-NEXT:    v_readfirstlane_b32 s8, v28
; GFX9-NEXT:    v_readfirstlane_b32 s9, v27
; GFX9-NEXT:    v_readfirstlane_b32 s10, v26
; GFX9-NEXT:    v_readfirstlane_b32 s11, v25
; GFX9-NEXT:    v_readfirstlane_b32 s12, v24
; GFX9-NEXT:    v_readfirstlane_b32 s13, v23
; GFX9-NEXT:    v_readfirstlane_b32 s14, v22
; GFX9-NEXT:    v_readfirstlane_b32 s15, v21
; GFX9-NEXT:    v_readfirstlane_b32 s40, v20
; GFX9-NEXT:    v_readfirstlane_b32 s41, v19
; GFX9-NEXT:    v_readfirstlane_b32 s42, v18
; GFX9-NEXT:    v_readfirstlane_b32 s43, v17
; GFX9-NEXT:    v_readfirstlane_b32 s44, v16
; GFX9-NEXT:    v_readfirstlane_b32 s56, v12
; GFX9-NEXT:    v_readfirstlane_b32 s57, v11
; GFX9-NEXT:    v_readfirstlane_b32 s58, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s60, v8
; GFX9-NEXT:    v_readfirstlane_b32 s61, v7
; GFX9-NEXT:    v_readfirstlane_b32 s62, v6
; GFX9-NEXT:    v_readfirstlane_b32 s63, v5
; GFX9-NEXT:    v_readfirstlane_b32 s72, v4
; GFX9-NEXT:    v_readfirstlane_b32 s73, v3
; GFX9-NEXT:    v_readfirstlane_b32 s74, v2
; GFX9-NEXT:    v_readfirstlane_b32 s75, v1
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    s_cbranch_scc0 .LBB59_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v12
; GFX9-NEXT:    v_perm_b32 v12, v14, v13, s4
; GFX9-NEXT:    v_perm_b32 v11, s6, v15, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_perm_b32 v12, v52, v51, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v54, v53, s4
; GFX9-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX9-NEXT:    v_perm_b32 v13, v48, v39, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v50, v49, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX9-NEXT:    v_perm_b32 v14, v36, v35, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v38, v37, s4
; GFX9-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX9-NEXT:    v_perm_b32 v15, v32, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    v_perm_b32 v16, v34, v33, s4
; GFX9-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX9-NEXT:    v_perm_b32 v16, v60, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_perm_b32 v17, v62, v61, s4
; GFX9-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX9-NEXT:    v_perm_b32 v17, v59, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_perm_b32 v18, v41, v40, s4
; GFX9-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX9-NEXT:    v_perm_b32 v18, v57, v19, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_perm_b32 v19, v42, v58, s4
; GFX9-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX9-NEXT:    v_perm_b32 v19, v45, v20, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX9-NEXT:    v_perm_b32 v20, v56, v47, s4
; GFX9-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX9-NEXT:    v_perm_b32 v20, v44, v21, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX9-NEXT:    v_perm_b32 v21, v46, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX9-NEXT:    v_perm_b32 v21, v31, v22, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX9-NEXT:    v_perm_b32 v22, v24, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX9-NEXT:    v_perm_b32 v22, v28, v23, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX9-NEXT:    v_perm_b32 v23, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX9-NEXT:    v_perm_b32 v23, v25, v63, s4
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX9-NEXT:    v_perm_b32 v24, v27, v55, s4
; GFX9-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s4
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    v_or_b32_e32 v30, v31, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v43, v46, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v31, v43, v31
; GFX9-NEXT:    s_mov_b64 s[4:5], 0
; GFX9-NEXT:    s_branch .LBB59_3
; GFX9-NEXT:  .LBB59_2:
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    s_mov_b64 s[4:5], -1
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:  .LBB59_3: ; %Flow
; GFX9-NEXT:    v_mov_b32_e32 v43, v44
; GFX9-NEXT:    v_mov_b32_e32 v44, v45
; GFX9-NEXT:    v_mov_b32_e32 v45, v47
; GFX9-NEXT:    v_mov_b32_e32 v46, v57
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_vccnz .LBB59_5
; GFX9-NEXT:  ; %bb.4: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    s_movk_i32 s4, 0x300
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    s_add_i32 s76, s76, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_add_u32_e32 v3, 0x300, v3
; GFX9-NEXT:    v_add_u32_sdwa v4, v4, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s74, s74, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    s_add_i32 s72, s72, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_add_u32_e32 v4, 0x300, v4
; GFX9-NEXT:    v_add_u32_sdwa v5, v5, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s62, s62, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    s_add_i32 s60, s60, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_add_u32_e32 v5, 0x300, v5
; GFX9-NEXT:    v_add_u32_sdwa v6, v6, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s58, s58, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    s_add_i32 s56, s56, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_add_u32_e32 v6, 0x300, v6
; GFX9-NEXT:    v_add_u32_sdwa v7, v7, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_add_u32_e32 v7, 0x300, v7
; GFX9-NEXT:    v_add_u32_sdwa v8, v8, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_add_u32_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u32_sdwa v9, v9, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_add_u32_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u32_sdwa v10, v10, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s5, 0xc0c0004
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v32
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v48
; GFX9-NEXT:    v_perm_b32 v14, v14, v39, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v14, v14, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v36
; GFX9-NEXT:    v_perm_b32 v15, v15, v35, s5
; GFX9-NEXT:    v_add_u32_sdwa v15, v15, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v60
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_perm_b32 v17, v17, v18, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v57
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v18, v18, v19, s5
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v11, s6, v12, v11
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u32_sdwa v16, v16, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v17, v17, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v18, v18, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_perm_b32 v12, v12, v13, s5
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v54
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v52
; GFX9-NEXT:    v_perm_b32 v12, v12, v53, s5
; GFX9-NEXT:    v_perm_b32 v13, v13, v51, s5
; GFX9-NEXT:    v_add_u32_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u32_sdwa v13, v13, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v50
; GFX9-NEXT:    v_perm_b32 v13, v13, v49, s5
; GFX9-NEXT:    v_add_u32_e32 v13, 0x300, v13
; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v38
; GFX9-NEXT:    v_perm_b32 v14, v14, v37, s5
; GFX9-NEXT:    v_add_u32_e32 v14, 0x300, v14
; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v34
; GFX9-NEXT:    v_perm_b32 v15, v15, v33, s5
; GFX9-NEXT:    v_add_u32_e32 v15, 0x300, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v62
; GFX9-NEXT:    v_perm_b32 v16, v16, v61, s5
; GFX9-NEXT:    v_add_u32_e32 v16, 0x300, v16
; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v41
; GFX9-NEXT:    v_perm_b32 v17, v17, v40, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 0x300, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v42
; GFX9-NEXT:    v_perm_b32 v18, v18, v47, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 0x300, v18
; GFX9-NEXT:    v_add_u32_sdwa v19, v19, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v46
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_perm_b32 v20, v20, v21, s5
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v20, v20, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v45
; GFX9-NEXT:    v_perm_b32 v20, v20, v44, s5
; GFX9-NEXT:    v_add_u32_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v21, v21, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    v_add_u32_sdwa v22, v22, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    v_add_u32_sdwa v23, v23, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    v_add_u32_sdwa v24, v24, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    v_add_u32_sdwa v25, v25, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    v_add_u32_sdwa v26, v26, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    v_add_u32_sdwa v27, v27, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    v_add_u32_sdwa v28, v28, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    v_add_u32_sdwa v29, v29, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    v_add_u32_sdwa v30, v30, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    v_add_u32_sdwa v31, v31, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v32, 3, v32
; GFX9-NEXT:    v_perm_b32 v32, v32, v33, s5
; GFX9-NEXT:    v_add_u32_sdwa v32, v32, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:  .LBB59_5: ; %end
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v16i64_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:260
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:188
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:60
; GFX11-TRUE16-NEXT:    s_clause 0xf
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s75, 0
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB59_4
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v135, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v55, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v147, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v99, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v103, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v67, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB59_3
; GFX11-TRUE16-NEXT:  .LBB59_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v135
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v99
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v55
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-TRUE16-NEXT:  .LBB59_3: ; %end
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT:  .LBB59_4:
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_branch .LBB59_2
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v16i64_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v32, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v34, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v33, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v37, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v35, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v39, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v36, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v48, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v38, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v49, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v50, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v52, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v51, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:260
; GFX11-FAKE16-NEXT:    scratch_load_u16 v53, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v54, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v55, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:188
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:132
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:60
; GFX11-FAKE16-NEXT:    s_clause 0xf
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s75, 0
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-FAKE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB59_4
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v144, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v64, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v147, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v103, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v67, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB59_3
; GFX11-FAKE16-NEXT:  .LBB59_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v144
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v100
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v64
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-FAKE16-NEXT:  .LBB59_3: ; %end
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-FAKE16-NEXT:  .LBB59_4:
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_branch .LBB59_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define <64 x bfloat> @bitcast_v16i64_to_v64bf16(<16 x i64> %a, i32 %b) {
; SI-LABEL: bitcast_v16i64_to_v64bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; kill: killed $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB60_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v62
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v62
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v13
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v11
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v10
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v9
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v8
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v7
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v3
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v2
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:  .LBB60_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB60_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v62, vcc
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v13
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v12
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v11
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v10
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v9
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v8
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v7
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v6
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v3
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v2
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v1
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v0
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:  .LBB60_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v60
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v58
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v56
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v46
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v44
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v42
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v40
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v54
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v52
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v50
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v48
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v39
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v38
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v37
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v36
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v35
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v34
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v33
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v32
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_alignbit_b32 v18, v18, v19, 16
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_alignbit_b32 v19, v19, v20, 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_alignbit_b32 v21, v21, v22, 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_alignbit_b32 v22, v22, v23, 16
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_alignbit_b32 v23, v23, v24, 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_alignbit_b32 v24, v24, v25, 16
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_alignbit_b32 v25, v25, v26, 16
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_alignbit_b32 v26, v26, v27, 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_alignbit_b32 v27, v27, v28, 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_alignbit_b32 v28, v28, v29, 16
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_alignbit_b32 v29, v29, v30, 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_alignbit_b32 v30, v30, v31, 16
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v63
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16i64_to_v64bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB60_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; VI-NEXT:  .LBB60_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16i64_to_v64bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB60_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v14
; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v12
; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v10
; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v8
; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v6
; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v4
; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v2
; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v0
; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v28
; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v26
; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v24
; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v22
; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v20
; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v18, vcc, 3, v18
; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v16
; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
; GFX9-NEXT:  .LBB60_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16i64_to_v64bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB60_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 3
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
; GFX11-NEXT:  .LBB60_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define inreg <64 x bfloat> @bitcast_v16i64_to_v64bf16_scalar(<16 x i64> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16i64_to_v64bf16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v33, s30, 0
; SI-NEXT:    v_writelane_b32 v33, s31, 1
; SI-NEXT:    v_writelane_b32 v33, s34, 2
; SI-NEXT:    v_writelane_b32 v33, s35, 3
; SI-NEXT:    v_writelane_b32 v33, s36, 4
; SI-NEXT:    v_writelane_b32 v33, s37, 5
; SI-NEXT:    v_writelane_b32 v33, s38, 6
; SI-NEXT:    v_writelane_b32 v33, s39, 7
; SI-NEXT:    v_writelane_b32 v33, s48, 8
; SI-NEXT:    v_writelane_b32 v33, s49, 9
; SI-NEXT:    v_writelane_b32 v33, s50, 10
; SI-NEXT:    v_writelane_b32 v33, s51, 11
; SI-NEXT:    v_writelane_b32 v33, s52, 12
; SI-NEXT:    v_writelane_b32 v33, s53, 13
; SI-NEXT:    v_writelane_b32 v33, s54, 14
; SI-NEXT:    v_writelane_b32 v33, s55, 15
; SI-NEXT:    v_writelane_b32 v33, s64, 16
; SI-NEXT:    v_writelane_b32 v33, s65, 17
; SI-NEXT:    v_writelane_b32 v33, s66, 18
; SI-NEXT:    v_writelane_b32 v33, s67, 19
; SI-NEXT:    v_writelane_b32 v33, s68, 20
; SI-NEXT:    v_writelane_b32 v33, s69, 21
; SI-NEXT:    v_writelane_b32 v33, s70, 22
; SI-NEXT:    v_writelane_b32 v33, s71, 23
; SI-NEXT:    v_writelane_b32 v33, s80, 24
; SI-NEXT:    v_writelane_b32 v33, s81, 25
; SI-NEXT:    v_writelane_b32 v33, s82, 26
; SI-NEXT:    v_writelane_b32 v33, s83, 27
; SI-NEXT:    v_writelane_b32 v33, s84, 28
; SI-NEXT:    v_writelane_b32 v33, s85, 29
; SI-NEXT:    v_writelane_b32 v33, s86, 30
; SI-NEXT:    v_writelane_b32 v33, s87, 31
; SI-NEXT:    v_writelane_b32 v33, s96, 32
; SI-NEXT:    v_writelane_b32 v33, s97, 33
; SI-NEXT:    v_writelane_b32 v33, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v33, s99, 35
; SI-NEXT:    v_readfirstlane_b32 s70, v17
; SI-NEXT:    v_readfirstlane_b32 s71, v16
; SI-NEXT:    v_readfirstlane_b32 s80, v15
; SI-NEXT:    v_readfirstlane_b32 s81, v14
; SI-NEXT:    v_readfirstlane_b32 s82, v13
; SI-NEXT:    v_readfirstlane_b32 s83, v12
; SI-NEXT:    v_readfirstlane_b32 s84, v11
; SI-NEXT:    v_readfirstlane_b32 s85, v10
; SI-NEXT:    v_readfirstlane_b32 s86, v9
; SI-NEXT:    v_readfirstlane_b32 s87, v8
; SI-NEXT:    v_readfirstlane_b32 s96, v7
; SI-NEXT:    v_readfirstlane_b32 s97, v6
; SI-NEXT:    v_readfirstlane_b32 s98, v5
; SI-NEXT:    v_readfirstlane_b32 s99, v4
; SI-NEXT:    v_readfirstlane_b32 s6, v3
; SI-NEXT:    v_readfirstlane_b32 s7, v2
; SI-NEXT:    v_readfirstlane_b32 s8, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s9, v0
; SI-NEXT:    ; implicit-def: $vgpr34 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB61_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s70, 0xffff0000
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v34, s4, 0
; SI-NEXT:    s_lshl_b32 s4, s70, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 1
; SI-NEXT:    s_and_b32 s4, s71, 0xffff0000
; SI-NEXT:    v_writelane_b32 v34, s4, 2
; SI-NEXT:    s_lshl_b32 s4, s71, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 3
; SI-NEXT:    s_and_b32 s11, s80, 0xffff0000
; SI-NEXT:    s_lshl_b32 s10, s80, 16
; SI-NEXT:    s_and_b32 s13, s81, 0xffff0000
; SI-NEXT:    s_lshl_b32 s12, s81, 16
; SI-NEXT:    s_and_b32 s15, s82, 0xffff0000
; SI-NEXT:    s_lshl_b32 s14, s82, 16
; SI-NEXT:    s_and_b32 s41, s83, 0xffff0000
; SI-NEXT:    s_lshl_b32 s40, s83, 16
; SI-NEXT:    s_and_b32 s43, s84, 0xffff0000
; SI-NEXT:    s_lshl_b32 s42, s84, 16
; SI-NEXT:    s_and_b32 s45, s85, 0xffff0000
; SI-NEXT:    s_lshl_b32 s44, s85, 16
; SI-NEXT:    s_and_b32 s47, s86, 0xffff0000
; SI-NEXT:    s_lshl_b32 s46, s86, 16
; SI-NEXT:    s_and_b32 s57, s87, 0xffff0000
; SI-NEXT:    s_lshl_b32 s56, s87, 16
; SI-NEXT:    s_and_b32 s59, s96, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s96, 16
; SI-NEXT:    s_and_b32 s61, s97, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s97, 16
; SI-NEXT:    s_and_b32 s63, s98, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s98, 16
; SI-NEXT:    s_and_b32 s73, s99, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s99, 16
; SI-NEXT:    s_and_b32 s75, s6, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s6, 16
; SI-NEXT:    s_and_b32 s77, s7, 0xffff0000
; SI-NEXT:    s_lshl_b32 s76, s7, 16
; SI-NEXT:    s_and_b32 s79, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s78, s8, 16
; SI-NEXT:    s_and_b32 s89, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s9, 16
; SI-NEXT:    s_and_b32 s91, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s29, 16
; SI-NEXT:    s_and_b32 s93, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s28, 16
; SI-NEXT:    s_and_b32 s95, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s27, 16
; SI-NEXT:    s_and_b32 s31, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s30, s26, 16
; SI-NEXT:    s_and_b32 s35, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s34, s25, 16
; SI-NEXT:    s_and_b32 s37, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s36, s24, 16
; SI-NEXT:    s_and_b32 s39, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s38, s23, 16
; SI-NEXT:    s_and_b32 s49, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s48, s22, 16
; SI-NEXT:    s_and_b32 s51, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s50, s21, 16
; SI-NEXT:    s_and_b32 s53, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s52, s20, 16
; SI-NEXT:    s_and_b32 s55, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s54, s19, 16
; SI-NEXT:    s_and_b32 s65, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s64, s18, 16
; SI-NEXT:    s_and_b32 s67, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s66, s17, 16
; SI-NEXT:    s_and_b32 s69, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s68, s16, 16
; SI-NEXT:    s_cbranch_execnz .LBB61_3
; SI-NEXT:  .LBB61_2: ; %cmp.true
; SI-NEXT:    s_add_u32 s4, s16, 3
; SI-NEXT:    s_addc_u32 s5, s17, 0
; SI-NEXT:    s_add_u32 s16, s18, 3
; SI-NEXT:    s_addc_u32 s17, s19, 0
; SI-NEXT:    s_add_u32 s18, s20, 3
; SI-NEXT:    s_addc_u32 s19, s21, 0
; SI-NEXT:    s_add_u32 s20, s22, 3
; SI-NEXT:    s_addc_u32 s21, s23, 0
; SI-NEXT:    s_add_u32 s22, s24, 3
; SI-NEXT:    s_addc_u32 s23, s25, 0
; SI-NEXT:    s_add_u32 s24, s26, 3
; SI-NEXT:    s_addc_u32 s25, s27, 0
; SI-NEXT:    s_add_u32 s26, s28, 3
; SI-NEXT:    s_addc_u32 s27, s29, 0
; SI-NEXT:    s_add_u32 s9, s9, 3
; SI-NEXT:    s_addc_u32 s8, s8, 0
; SI-NEXT:    s_add_u32 s7, s7, 3
; SI-NEXT:    s_addc_u32 s6, s6, 0
; SI-NEXT:    s_add_u32 s28, s99, 3
; SI-NEXT:    s_addc_u32 s29, s98, 0
; SI-NEXT:    s_add_u32 s60, s97, 3
; SI-NEXT:    s_addc_u32 s58, s96, 0
; SI-NEXT:    s_add_u32 s56, s87, 3
; SI-NEXT:    s_addc_u32 s46, s86, 0
; SI-NEXT:    s_add_u32 s44, s85, 3
; SI-NEXT:    s_addc_u32 s42, s84, 0
; SI-NEXT:    s_add_u32 s40, s83, 3
; SI-NEXT:    s_addc_u32 s14, s82, 0
; SI-NEXT:    s_add_u32 s12, s81, 3
; SI-NEXT:    s_addc_u32 s10, s80, 0
; SI-NEXT:    s_add_u32 s11, s71, 3
; SI-NEXT:    s_addc_u32 s13, s70, 0
; SI-NEXT:    s_and_b32 s15, s13, 0xffff0000
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v34, s15, 0
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    v_writelane_b32 v34, s13, 1
; SI-NEXT:    s_and_b32 s13, s11, 0xffff0000
; SI-NEXT:    v_writelane_b32 v34, s13, 2
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    v_writelane_b32 v34, s11, 3
; SI-NEXT:    s_and_b32 s11, s10, 0xffff0000
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_and_b32 s13, s12, 0xffff0000
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_and_b32 s15, s14, 0xffff0000
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_and_b32 s41, s40, 0xffff0000
; SI-NEXT:    s_lshl_b32 s40, s40, 16
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_and_b32 s63, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s29, 16
; SI-NEXT:    s_and_b32 s73, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s28, 16
; SI-NEXT:    s_and_b32 s75, s6, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s6, 16
; SI-NEXT:    s_and_b32 s77, s7, 0xffff0000
; SI-NEXT:    s_lshl_b32 s76, s7, 16
; SI-NEXT:    s_and_b32 s79, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s78, s8, 16
; SI-NEXT:    s_and_b32 s89, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s9, 16
; SI-NEXT:    s_and_b32 s91, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s27, 16
; SI-NEXT:    s_and_b32 s93, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s26, 16
; SI-NEXT:    s_and_b32 s95, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s25, 16
; SI-NEXT:    s_and_b32 s31, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s30, s24, 16
; SI-NEXT:    s_and_b32 s35, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s34, s23, 16
; SI-NEXT:    s_and_b32 s37, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s36, s22, 16
; SI-NEXT:    s_and_b32 s39, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s38, s21, 16
; SI-NEXT:    s_and_b32 s49, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s48, s20, 16
; SI-NEXT:    s_and_b32 s51, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s50, s19, 16
; SI-NEXT:    s_and_b32 s53, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s52, s18, 16
; SI-NEXT:    s_and_b32 s55, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s54, s17, 16
; SI-NEXT:    s_and_b32 s65, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s64, s16, 16
; SI-NEXT:    s_and_b32 s67, s5, 0xffff0000
; SI-NEXT:    s_lshl_b32 s66, s5, 16
; SI-NEXT:    s_and_b32 s69, s4, 0xffff0000
; SI-NEXT:    s_lshl_b32 s68, s4, 16
; SI-NEXT:  .LBB61_3: ; %end
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s69
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s68
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s67
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s66
; SI-NEXT:    v_lshr_b64 v[1:2], v[1:2], 16
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s65
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s64
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s55
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s54
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s53
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s52
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s51
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s50
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s49
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s48
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s39
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s38
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s37
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s36
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s35
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s34
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s31
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s30
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s95
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s94
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s93
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s92
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s91
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s90
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s89
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s88
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s79
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s78
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s77
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s76
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s75
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s74
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s73
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s72
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s63
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s62
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s61
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s60
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s59
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s58
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s57
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s56
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s47
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s46
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s45
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s44
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s43
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s42
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s41
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v26
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s40
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s15
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v27
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s14
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s13
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v28
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s12
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s11
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v29
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s10
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_readlane_b32 s4, v34, 2
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s4
; SI-NEXT:    v_readlane_b32 s4, v34, 3
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v30
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s4
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_readlane_b32 s4, v34, 0
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s4
; SI-NEXT:    v_readlane_b32 s4, v34, 1
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v31
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s4
; SI-NEXT:    v_lshr_b64 v[31:32], v[31:32], 16
; SI-NEXT:    v_readlane_b32 s99, v33, 35
; SI-NEXT:    v_readlane_b32 s98, v33, 34
; SI-NEXT:    v_readlane_b32 s97, v33, 33
; SI-NEXT:    v_readlane_b32 s96, v33, 32
; SI-NEXT:    v_readlane_b32 s87, v33, 31
; SI-NEXT:    v_readlane_b32 s86, v33, 30
; SI-NEXT:    v_readlane_b32 s85, v33, 29
; SI-NEXT:    v_readlane_b32 s84, v33, 28
; SI-NEXT:    v_readlane_b32 s83, v33, 27
; SI-NEXT:    v_readlane_b32 s82, v33, 26
; SI-NEXT:    v_readlane_b32 s81, v33, 25
; SI-NEXT:    v_readlane_b32 s80, v33, 24
; SI-NEXT:    v_readlane_b32 s71, v33, 23
; SI-NEXT:    v_readlane_b32 s70, v33, 22
; SI-NEXT:    v_readlane_b32 s69, v33, 21
; SI-NEXT:    v_readlane_b32 s68, v33, 20
; SI-NEXT:    v_readlane_b32 s67, v33, 19
; SI-NEXT:    v_readlane_b32 s66, v33, 18
; SI-NEXT:    v_readlane_b32 s65, v33, 17
; SI-NEXT:    v_readlane_b32 s64, v33, 16
; SI-NEXT:    v_readlane_b32 s55, v33, 15
; SI-NEXT:    v_readlane_b32 s54, v33, 14
; SI-NEXT:    v_readlane_b32 s53, v33, 13
; SI-NEXT:    v_readlane_b32 s52, v33, 12
; SI-NEXT:    v_readlane_b32 s51, v33, 11
; SI-NEXT:    v_readlane_b32 s50, v33, 10
; SI-NEXT:    v_readlane_b32 s49, v33, 9
; SI-NEXT:    v_readlane_b32 s48, v33, 8
; SI-NEXT:    v_readlane_b32 s39, v33, 7
; SI-NEXT:    v_readlane_b32 s38, v33, 6
; SI-NEXT:    v_readlane_b32 s37, v33, 5
; SI-NEXT:    v_readlane_b32 s36, v33, 4
; SI-NEXT:    v_readlane_b32 s35, v33, 3
; SI-NEXT:    v_readlane_b32 s34, v33, 2
; SI-NEXT:    v_readlane_b32 s31, v33, 1
; SI-NEXT:    v_readlane_b32 s30, v33, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB61_4:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr37
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr35
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr31
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr93
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr77
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr75
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr73
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr63
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr61
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr59
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr57
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr47
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr45
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    ; implicit-def: $sgpr43
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr41
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr15
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr13
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr11
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    s_branch .LBB61_2
;
; VI-LABEL: bitcast_v16i64_to_v64bf16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB61_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB61_3
; VI-NEXT:  .LBB61_2: ; %cmp.true
; VI-NEXT:    s_add_u32 s47, s47, 3
; VI-NEXT:    s_addc_u32 s46, s46, 0
; VI-NEXT:    s_add_u32 s28, s28, 3
; VI-NEXT:    s_addc_u32 s29, s29, 0
; VI-NEXT:    s_add_u32 s26, s26, 3
; VI-NEXT:    s_addc_u32 s27, s27, 0
; VI-NEXT:    s_add_u32 s24, s24, 3
; VI-NEXT:    s_addc_u32 s25, s25, 0
; VI-NEXT:    s_add_u32 s22, s22, 3
; VI-NEXT:    s_addc_u32 s23, s23, 0
; VI-NEXT:    s_add_u32 s20, s20, 3
; VI-NEXT:    s_addc_u32 s21, s21, 0
; VI-NEXT:    s_add_u32 s18, s18, 3
; VI-NEXT:    s_addc_u32 s19, s19, 0
; VI-NEXT:    s_add_u32 s16, s16, 3
; VI-NEXT:    s_addc_u32 s17, s17, 0
; VI-NEXT:    s_add_u32 s7, s7, 3
; VI-NEXT:    s_addc_u32 s6, s6, 0
; VI-NEXT:    s_add_u32 s9, s9, 3
; VI-NEXT:    s_addc_u32 s8, s8, 0
; VI-NEXT:    s_add_u32 s11, s11, 3
; VI-NEXT:    s_addc_u32 s10, s10, 0
; VI-NEXT:    s_add_u32 s13, s13, 3
; VI-NEXT:    s_addc_u32 s12, s12, 0
; VI-NEXT:    s_add_u32 s15, s15, 3
; VI-NEXT:    s_addc_u32 s14, s14, 0
; VI-NEXT:    s_add_u32 s41, s41, 3
; VI-NEXT:    s_addc_u32 s40, s40, 0
; VI-NEXT:    s_add_u32 s43, s43, 3
; VI-NEXT:    s_addc_u32 s42, s42, 0
; VI-NEXT:    s_add_u32 s45, s45, 3
; VI-NEXT:    s_addc_u32 s44, s44, 0
; VI-NEXT:  .LBB61_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB61_4:
; VI-NEXT:    s_branch .LBB61_2
;
; GFX9-LABEL: bitcast_v16i64_to_v64bf16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB61_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB61_3
; GFX9-NEXT:  .LBB61_2: ; %cmp.true
; GFX9-NEXT:    s_add_u32 s47, s47, 3
; GFX9-NEXT:    s_addc_u32 s46, s46, 0
; GFX9-NEXT:    s_add_u32 s28, s28, 3
; GFX9-NEXT:    s_addc_u32 s29, s29, 0
; GFX9-NEXT:    s_add_u32 s26, s26, 3
; GFX9-NEXT:    s_addc_u32 s27, s27, 0
; GFX9-NEXT:    s_add_u32 s24, s24, 3
; GFX9-NEXT:    s_addc_u32 s25, s25, 0
; GFX9-NEXT:    s_add_u32 s22, s22, 3
; GFX9-NEXT:    s_addc_u32 s23, s23, 0
; GFX9-NEXT:    s_add_u32 s20, s20, 3
; GFX9-NEXT:    s_addc_u32 s21, s21, 0
; GFX9-NEXT:    s_add_u32 s18, s18, 3
; GFX9-NEXT:    s_addc_u32 s19, s19, 0
; GFX9-NEXT:    s_add_u32 s16, s16, 3
; GFX9-NEXT:    s_addc_u32 s17, s17, 0
; GFX9-NEXT:    s_add_u32 s7, s7, 3
; GFX9-NEXT:    s_addc_u32 s6, s6, 0
; GFX9-NEXT:    s_add_u32 s9, s9, 3
; GFX9-NEXT:    s_addc_u32 s8, s8, 0
; GFX9-NEXT:    s_add_u32 s11, s11, 3
; GFX9-NEXT:    s_addc_u32 s10, s10, 0
; GFX9-NEXT:    s_add_u32 s13, s13, 3
; GFX9-NEXT:    s_addc_u32 s12, s12, 0
; GFX9-NEXT:    s_add_u32 s15, s15, 3
; GFX9-NEXT:    s_addc_u32 s14, s14, 0
; GFX9-NEXT:    s_add_u32 s41, s41, 3
; GFX9-NEXT:    s_addc_u32 s40, s40, 0
; GFX9-NEXT:    s_add_u32 s43, s43, 3
; GFX9-NEXT:    s_addc_u32 s42, s42, 0
; GFX9-NEXT:    s_add_u32 s45, s45, 3
; GFX9-NEXT:    s_addc_u32 s44, s44, 0
; GFX9-NEXT:  .LBB61_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB61_4:
; GFX9-NEXT:    s_branch .LBB61_2
;
; GFX11-LABEL: bitcast_v16i64_to_v64bf16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB61_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB61_3
; GFX11-NEXT:  .LBB61_2: ; %cmp.true
; GFX11-NEXT:    s_add_u32 s26, s26, 3
; GFX11-NEXT:    s_addc_u32 s27, s27, 0
; GFX11-NEXT:    s_add_u32 s24, s24, 3
; GFX11-NEXT:    s_addc_u32 s25, s25, 0
; GFX11-NEXT:    s_add_u32 s22, s22, 3
; GFX11-NEXT:    s_addc_u32 s23, s23, 0
; GFX11-NEXT:    s_add_u32 s20, s20, 3
; GFX11-NEXT:    s_addc_u32 s21, s21, 0
; GFX11-NEXT:    s_add_u32 s18, s18, 3
; GFX11-NEXT:    s_addc_u32 s19, s19, 0
; GFX11-NEXT:    s_add_u32 s16, s16, 3
; GFX11-NEXT:    s_addc_u32 s17, s17, 0
; GFX11-NEXT:    s_add_u32 s2, s2, 3
; GFX11-NEXT:    s_addc_u32 s3, s3, 0
; GFX11-NEXT:    s_add_u32 s0, s0, 3
; GFX11-NEXT:    s_addc_u32 s1, s1, 0
; GFX11-NEXT:    s_add_u32 s5, s5, 3
; GFX11-NEXT:    s_addc_u32 s4, s4, 0
; GFX11-NEXT:    s_add_u32 s7, s7, 3
; GFX11-NEXT:    s_addc_u32 s6, s6, 0
; GFX11-NEXT:    s_add_u32 s9, s9, 3
; GFX11-NEXT:    s_addc_u32 s8, s8, 0
; GFX11-NEXT:    s_add_u32 s11, s11, 3
; GFX11-NEXT:    s_addc_u32 s10, s10, 0
; GFX11-NEXT:    s_add_u32 s13, s13, 3
; GFX11-NEXT:    s_addc_u32 s12, s12, 0
; GFX11-NEXT:    s_add_u32 s15, s15, 3
; GFX11-NEXT:    s_addc_u32 s14, s14, 0
; GFX11-NEXT:    s_add_u32 s41, s41, 3
; GFX11-NEXT:    s_addc_u32 s40, s40, 0
; GFX11-NEXT:    s_add_u32 s28, s28, 3
; GFX11-NEXT:    s_addc_u32 s29, s29, 0
; GFX11-NEXT:  .LBB61_3: ; %end
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB61_4:
; GFX11-NEXT:    s_branch .LBB61_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define <16 x i64> @bitcast_v64bf16_to_v16i64(<64 x bfloat> %a, i32 %b) {
; SI-LABEL: bitcast_v64bf16_to_v16i64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v0
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v63
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v62
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v60
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v58, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_and_b32_e32 v46, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v36, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v53
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v37
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v38
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v38
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v42
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v40
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v55
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v18
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v19
; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v51
; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v20
; SI-NEXT:    v_mul_f32_e32 v45, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v49
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v41, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v23
; SI-NEXT:    v_mul_f32_e32 v55, 1.0, v39
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v53, 1.0, v36
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v25
; SI-NEXT:    v_mul_f32_e32 v51, 1.0, v35
; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v49, 1.0, v34
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v39, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v29
; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v31
; SI-NEXT:    v_mul_f32_e32 v35, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v33, 1.0, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB62_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v51
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v39
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v36
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v34
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_alignbit_b32 v19, v19, v59, 16
; SI-NEXT:    v_alignbit_b32 v20, v20, v57, 16
; SI-NEXT:    v_alignbit_b32 v21, v21, v46, 16
; SI-NEXT:    v_alignbit_b32 v22, v22, v44, 16
; SI-NEXT:    v_alignbit_b32 v23, v23, v42, 16
; SI-NEXT:    v_alignbit_b32 v24, v24, v40, 16
; SI-NEXT:    v_alignbit_b32 v25, v25, v54, 16
; SI-NEXT:    v_alignbit_b32 v26, v26, v52, 16
; SI-NEXT:    v_alignbit_b32 v27, v27, v50, 16
; SI-NEXT:    v_alignbit_b32 v28, v28, v48, 16
; SI-NEXT:    v_alignbit_b32 v29, v29, v38, 16
; SI-NEXT:    v_alignbit_b32 v30, v30, v35, 16
; SI-NEXT:    v_alignbit_b32 v31, v31, v33, 16
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_alignbit_b32 v0, v0, v63, 16
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v62
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v60
; SI-NEXT:    v_alignbit_b32 v18, v18, v61, 16
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:  .LBB62_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB62_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v63
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v60
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v55
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v0, v1, v0, 16
; SI-NEXT:    v_alignbit_b32 v1, v3, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v2, v3, v2, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_alignbit_b32 v3, v4, v3, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_alignbit_b32 v4, v5, v4, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_alignbit_b32 v5, v6, v5, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_alignbit_b32 v6, v7, v6, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_alignbit_b32 v7, v8, v7, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_alignbit_b32 v8, v9, v8, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_alignbit_b32 v9, v10, v9, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_alignbit_b32 v10, v11, v10, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_alignbit_b32 v11, v12, v11, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_alignbit_b32 v12, v13, v12, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v13, v14, v13, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_alignbit_b32 v14, v15, v14, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v15, v16, v15, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v16, v17, v16, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_alignbit_b32 v17, v18, v17, 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v61
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_alignbit_b32 v18, v19, v18, 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v59
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_alignbit_b32 v19, v20, v19, 16
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v57
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_alignbit_b32 v20, v21, v20, 16
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_alignbit_b32 v21, v22, v21, 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v44
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_alignbit_b32 v22, v23, v22, 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_alignbit_b32 v23, v24, v23, 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_alignbit_b32 v24, v25, v24, 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_alignbit_b32 v25, v26, v25, 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_alignbit_b32 v26, v27, v26, 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_alignbit_b32 v27, v28, v27, 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_alignbit_b32 v28, v29, v28, 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_alignbit_b32 v29, v30, v29, 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v35
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_alignbit_b32 v30, v31, v30, 16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v33
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_alignbit_b32 v31, v32, v31, 16
; SI-NEXT:  .LBB62_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v16i64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB62_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v15, 16, 1
; VI-NEXT:    s_movk_i32 s6, 0x7fff
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v15
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_alignbit_b32 v15, v15, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v14, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v14
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; VI-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_alignbit_b32 v14, v14, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v13
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_alignbit_b32 v13, v13, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v12, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v12
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; VI-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_alignbit_b32 v12, v12, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v11
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_alignbit_b32 v11, v11, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v10, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v10
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; VI-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_alignbit_b32 v10, v10, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v9
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_alignbit_b32 v9, v9, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v8, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v8
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; VI-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_alignbit_b32 v8, v8, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v7
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_alignbit_b32 v7, v7, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v6, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v6
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; VI-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_alignbit_b32 v6, v6, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v5
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_alignbit_b32 v5, v5, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v4, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v4
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; VI-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_alignbit_b32 v4, v4, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v3
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_alignbit_b32 v3, v3, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v2
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_alignbit_b32 v2, v2, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v1
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_alignbit_b32 v1, v1, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v0
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; VI-NEXT:    v_alignbit_b32 v0, v0, v32, 16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v30, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v30
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; VI-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_alignbit_b32 v30, v30, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v29, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v29
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_alignbit_b32 v29, v29, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v28, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v28
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; VI-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_alignbit_b32 v28, v28, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v27
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_alignbit_b32 v27, v27, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v26, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v26
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; VI-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_alignbit_b32 v26, v26, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v25
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_alignbit_b32 v25, v25, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v24, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v24
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; VI-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_alignbit_b32 v24, v24, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v23
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_alignbit_b32 v23, v23, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v22, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v22
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; VI-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_alignbit_b32 v22, v22, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v21
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_alignbit_b32 v21, v21, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v20, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v20
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; VI-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_alignbit_b32 v20, v20, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v19
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_alignbit_b32 v19, v19, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v18
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_alignbit_b32 v18, v18, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v17
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_alignbit_b32 v17, v17, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v16
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_alignbit_b32 v16, v16, v32, 16
; VI-NEXT:  .LBB62_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v16i64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB62_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    s_movk_i32 s6, 0x7fff
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v15, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v15, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; GFX9-NEXT:    s_mov_b32 s7, 0x7060302
; GFX9-NEXT:    v_perm_b32 v15, v15, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v14, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v14, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v14, v14, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v13, v13, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v12, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v12, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v12, v12, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v11, v11, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v10, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v10, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v10, v10, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v9, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v9, v9, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v8, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v8, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v8, v8, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v7, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v7, v7, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v6, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v6, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v6, v6, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v5, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v5, v5, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v4, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v4, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v4, v4, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v3, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v3, v3, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v2, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v2, v2, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v1, v1, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v0, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v0, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v0, v0, v32, s7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v30, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v30, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v30, v30, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v29, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v29, v29, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v28, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v28, v28, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v27, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v27, v27, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v26, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v26, v26, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v25, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v25, v25, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v24, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v24, v24, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v23, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v23, v23, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v22, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v22, v22, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v21, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v21, v21, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v20, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v20, v20, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v19, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v19, v19, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v18, v18, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v17, v17, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v16, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v16, v16, v32, s7
; GFX9-NEXT:  .LBB62_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v16i64:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB62_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v63, v31 :: v_dual_mov_b32 v62, v30
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v61, v29 :: v_dual_mov_b32 v60, v28
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v59, v27 :: v_dual_mov_b32 v58, v26
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v57, v25 :: v_dual_mov_b32 v56, v24
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v23 :: v_dual_mov_b32 v54, v22
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v21 :: v_dual_mov_b32 v52, v20
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v19 :: v_dual_mov_b32 v50, v18
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v17 :: v_dual_mov_b32 v48, v16
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v47, v15 :: v_dual_mov_b32 v46, v14
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v45, v13 :: v_dual_mov_b32 v44, v12
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v43, v11 :: v_dual_mov_b32 v42, v10
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v41, v9 :: v_dual_mov_b32 v40, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v7 :: v_dual_mov_b32 v38, v6
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v5 :: v_dual_mov_b32 v36, v4
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v3 :: v_dual_mov_b32 v34, v2
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v1 :: v_dual_mov_b32 v32, v0
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT:  .LBB62_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB62_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v32, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v35, v38 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v39, v33, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v47, v36, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v32.h
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_add3_u32 v38, v48, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v38, v49, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v34 :: v_dual_cndmask_b32 v46, v35, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, v14.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v33, v34, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v13, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v45, v33, v35
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v13.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v15, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v44, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, v12.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v11, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v14, v32 :: v_dual_and_b32 v12, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v43, v15, v33
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v10, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v13, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v42, v14, v32, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v10.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v9, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v41, v13, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, v9.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v8, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v11, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v40, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v8.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v7, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v39, v11, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v7.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v9, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v6.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v8, v10 :: v_dual_and_b32 v6, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v37, v9, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v8, v10, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v4.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v6, v8 :: v_dual_and_b32 v4, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_cndmask_b32 v35, v7, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v5, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v6, v8, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v4, v6 :: v_dual_and_b32 v2, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_cndmask_b32 v33, v5, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v1.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v31
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v5, 16, v31
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v3, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v4, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v0.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v63, v3, v7 :: v_dual_and_b32 v0, 0xffff0000, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v62, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v28
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v61, v3, v7 :: v_dual_lshlrev_b32 v4, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v60, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v59, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v25
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v4, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v58, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v24
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v57, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v56, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v55, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v5, 16, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v53, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v52, v0, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v2.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_lshlrev_b32 v6, 16, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v6 :: v_dual_cndmask_b32 v51, v0, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v8, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v50, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v5, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v16
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v1 :: v_dual_add_f32 v1, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v0.h
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v9, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v1, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v5.h
; GFX11-TRUE16-NEXT:  .LBB62_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v40 :: v_dual_mov_b32 v9, v41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v42 :: v_dual_mov_b32 v11, v43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, v44 :: v_dual_mov_b32 v13, v45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v46 :: v_dual_mov_b32 v15, v47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v56 :: v_dual_mov_b32 v25, v57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v58 :: v_dual_mov_b32 v27, v59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v60 :: v_dual_mov_b32 v29, v61
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v62 :: v_dual_mov_b32 v31, v63
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v33
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v34 :: v_dual_mov_b32 v3, v35
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v36 :: v_dual_mov_b32 v5, v37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v38 :: v_dual_mov_b32 v7, v39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v48 :: v_dual_mov_b32 v17, v49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v50 :: v_dual_mov_b32 v19, v51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v52 :: v_dual_mov_b32 v21, v53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v54 :: v_dual_mov_b32 v23, v55
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v16i64:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB62_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v39, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v13, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v11, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v8, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v9, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v34 :: v_dual_add_f32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v6, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v6, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v7, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v5, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v3, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v0, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v1, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v31
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v31, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v29
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v30, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v29, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v28, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v27
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v28, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v27, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v26, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v25, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v23
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v23, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v36, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v21, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v32, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v38, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v20, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v19, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v48, v18, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v36, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v36
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v39, v17, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v16, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v35, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v37, v50, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v16, v36, 0x7060302
; GFX11-FAKE16-NEXT:  .LBB62_2: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define inreg <16 x i64> @bitcast_v64bf16_to_v16i64_scalar(<64 x bfloat> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64bf16_to_v16i64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    s_and_b32 s12, s25, 0xffff0000
; SI-NEXT:    s_and_b32 s30, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v1
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; SI-NEXT:    s_and_b32 s34, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_readfirstlane_b32 s94, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s34
; SI-NEXT:    v_readfirstlane_b32 s90, v5
; SI-NEXT:    s_and_b32 s95, s94, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s30
; SI-NEXT:    v_readfirstlane_b32 s88, v6
; SI-NEXT:    s_and_b32 s91, s90, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s95
; SI-NEXT:    v_readfirstlane_b32 s78, v7
; SI-NEXT:    s_and_b32 s89, s88, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s91
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    s_and_b32 s79, s78, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s89
; SI-NEXT:    v_readfirstlane_b32 s74, v9
; SI-NEXT:    s_and_b32 s77, s76, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s79
; SI-NEXT:    v_readfirstlane_b32 s72, v10
; SI-NEXT:    s_and_b32 s75, s74, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s77
; SI-NEXT:    v_readfirstlane_b32 s62, v11
; SI-NEXT:    s_and_b32 s73, s72, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s75
; SI-NEXT:    v_readfirstlane_b32 s60, v12
; SI-NEXT:    s_and_b32 s63, s62, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s73
; SI-NEXT:    v_readfirstlane_b32 s58, v13
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s63
; SI-NEXT:    v_readfirstlane_b32 s56, v14
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s61
; SI-NEXT:    v_readfirstlane_b32 s46, v15
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s59
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s57
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s47
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    v_readfirstlane_b32 s92, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s45
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s29, 16
; SI-NEXT:    s_lshl_b32 s7, s28, 16
; SI-NEXT:    s_and_b32 s8, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s9, s27, 16
; SI-NEXT:    s_and_b32 s10, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s11, s26, 16
; SI-NEXT:    s_lshl_b32 s13, s25, 16
; SI-NEXT:    s_and_b32 s14, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s15, s24, 16
; SI-NEXT:    s_and_b32 s24, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s25, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s26, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s27, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s28, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s29, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_and_b32 s41, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_and_b32 s93, s92, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_and_b32 s36, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s43
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cmp_lg_u32 vcc_lo, 0
; SI-NEXT:    v_mul_f32_e64 v55, 1.0, s41
; SI-NEXT:    v_mul_f32_e64 v53, 1.0, s40
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s29
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v62, 1.0, s27
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s26
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s25
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s24
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s14
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s10
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s8
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s4
; SI-NEXT:    v_mul_f32_e64 v39, 1.0, s36
; SI-NEXT:    v_mul_f32_e64 v47, 1.0, s93
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e64 v43, 1.0, s16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s17
; SI-NEXT:    v_mul_f32_e64 v34, 1.0, s18
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s19
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s20
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s21
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s22
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s15
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s13
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s11
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s9
; SI-NEXT:    v_mul_f32_e64 v61, 1.0, s7
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s5
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s37
; SI-NEXT:    v_mul_f32_e64 v60, 1.0, s35
; SI-NEXT:    v_mul_f32_e64 v33, 1.0, s31
; SI-NEXT:    v_mul_f32_e64 v57, 1.0, s94
; SI-NEXT:    v_mul_f32_e64 v59, 1.0, s92
; SI-NEXT:    v_mul_f32_e64 v56, 1.0, s90
; SI-NEXT:    v_mul_f32_e64 v38, 1.0, s88
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s78
; SI-NEXT:    v_mul_f32_e64 v36, 1.0, s76
; SI-NEXT:    v_mul_f32_e64 v46, 1.0, s74
; SI-NEXT:    v_mul_f32_e64 v48, 1.0, s72
; SI-NEXT:    v_mul_f32_e64 v49, 1.0, s62
; SI-NEXT:    v_mul_f32_e64 v45, 1.0, s60
; SI-NEXT:    v_mul_f32_e64 v42, 1.0, s58
; SI-NEXT:    v_mul_f32_e64 v51, 1.0, s56
; SI-NEXT:    v_mul_f32_e64 v41, 1.0, s46
; SI-NEXT:    v_mul_f32_e64 v54, 1.0, s44
; SI-NEXT:    v_mul_f32_e64 v52, 1.0, s42
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB63_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v55
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v53
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v44, v1
; SI-NEXT:    v_mov_b32_e32 v43, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v12
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    v_mov_b32_e32 v36, v47
; SI-NEXT:    v_mov_b32_e32 v50, v49
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v40, v41
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v1, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[43:44], 16
; SI-NEXT:    v_mov_b32_e32 v2, v34
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[2:3], v[34:35], 16
; SI-NEXT:    v_mov_b32_e32 v3, v17
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[3:4], v[17:18], 16
; SI-NEXT:    v_mov_b32_e32 v4, v20
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[4:5], v[20:21], 16
; SI-NEXT:    v_mov_b32_e32 v5, v19
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v6
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[5:6], v[19:20], 16
; SI-NEXT:    v_mov_b32_e32 v6, v13
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[6:7], v[13:14], 16
; SI-NEXT:    v_mov_b32_e32 v7, v31
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[7:8], v[31:32], 16
; SI-NEXT:    v_mov_b32_e32 v8, v23
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[8:9], v[23:24], 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v11
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v44, v45
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v9
; SI-NEXT:    v_mov_b32_e32 v9, v25
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[9:10], v[25:26], 16
; SI-NEXT:    v_mov_b32_e32 v10, v30
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v11, v27
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[11:12], v[27:28], 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v18
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v12
; SI-NEXT:    v_mov_b32_e32 v12, v61
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[12:13], v[61:62], 16
; SI-NEXT:    v_mov_b32_e32 v13, v29
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[13:14], v[29:30], 16
; SI-NEXT:    v_mov_b32_e32 v14, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v62, v39
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[14:15], v[15:16], 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v15
; SI-NEXT:    v_mov_b32_e32 v15, v60
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[15:16], v[60:61], 16
; SI-NEXT:    v_mov_b32_e32 v16, v33
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[16:17], v[33:34], 16
; SI-NEXT:    v_mov_b32_e32 v17, v57
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[17:18], v[57:58], 16
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v47
; SI-NEXT:    v_lshr_b64 v[18:19], v[59:60], 16
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v20
; SI-NEXT:    v_lshr_b64 v[19:20], v[56:57], 16
; SI-NEXT:    v_mov_b32_e32 v20, v38
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[20:21], v[38:39], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; SI-NEXT:    v_mov_b32_e32 v39, v62
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v21
; SI-NEXT:    v_mov_b32_e32 v21, v22
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[21:22], v[22:23], 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v55
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[37:38], 16
; SI-NEXT:    v_lshr_b64 v[23:24], v[46:47], 16
; SI-NEXT:    v_lshr_b64 v[24:25], v[48:49], 16
; SI-NEXT:    v_lshr_b64 v[25:26], v[50:51], 16
; SI-NEXT:    v_lshr_b64 v[26:27], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[27:28], v[42:43], 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_lshr_b64 v[61:62], v[52:53], 16
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v45, v44
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v47, v36
; SI-NEXT:    v_mov_b32_e32 v36, v37
; SI-NEXT:    v_mov_b32_e32 v49, v50
; SI-NEXT:    v_mov_b32_e32 v51, v33
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_lshr_b64 v[29:30], v[40:41], 16
; SI-NEXT:    v_lshr_b64 v[30:31], v[54:55], 16
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v31, v61
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    s_branch .LBB63_3
; SI-NEXT:  .LBB63_2:
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB63_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v37, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v33, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(3)
; SI-NEXT:    v_mov_b32_e32 v38, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v50, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v40, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_cbranch_vccnz .LBB63_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v55
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v34, v56
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshr_b64 v[1:2], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v33
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_lshr_b64 v[32:33], v[32:33], 16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v37
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v31, v32
; SI-NEXT:  .LBB63_5: ; %end
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v16i64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v35, s36, 0
; VI-NEXT:    v_writelane_b32 v35, s37, 1
; VI-NEXT:    v_writelane_b32 v35, s38, 2
; VI-NEXT:    v_writelane_b32 v35, s39, 3
; VI-NEXT:    v_writelane_b32 v35, s48, 4
; VI-NEXT:    v_writelane_b32 v35, s49, 5
; VI-NEXT:    v_writelane_b32 v35, s50, 6
; VI-NEXT:    v_writelane_b32 v35, s51, 7
; VI-NEXT:    v_writelane_b32 v35, s52, 8
; VI-NEXT:    v_writelane_b32 v35, s53, 9
; VI-NEXT:    v_writelane_b32 v35, s54, 10
; VI-NEXT:    v_writelane_b32 v35, s55, 11
; VI-NEXT:    v_writelane_b32 v35, s64, 12
; VI-NEXT:    v_writelane_b32 v35, s65, 13
; VI-NEXT:    v_writelane_b32 v35, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v35, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB63_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB63_4
; VI-NEXT:  .LBB63_2: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x40c00000
; VI-NEXT:    s_lshl_b32 s4, s50, 16
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    s_and_b32 s4, s50, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v2, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s48, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; VI-NEXT:    v_add_f32_e32 v2, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    s_and_b32 s4, s48, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v4, v3, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_lshrrev_b64 v[14:15], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s46, 16
; VI-NEXT:    v_lshrrev_b64 v[12:13], 16, v[2:3]
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[10:11], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[8:9], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s38, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[4:5], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s36, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[2:3], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; VI-NEXT:    v_bfe_u32 v3, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v0
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_lshrrev_b64 v[0:1], 16, v[0:1]
; VI-NEXT:    s_and_b32 s4, s66, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s66, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s64, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s64, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s62, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[30:31], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s62, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s60, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[28:29], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s60, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s58, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[26:27], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s58, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s56, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[24:25], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s56, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s54, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[22:23], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s54, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s52, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[20:21], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s52, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s37, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s37, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_and_b32 s4, s39, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v5, v7, vcc
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    s_lshl_b32 s4, s39, 16
; VI-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_mov_b32_e32 v1, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v7, v9, vcc
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    s_lshl_b32 s4, s41, 16
; VI-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v3
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_mov_b32_e32 v3, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v9, v11, vcc
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    s_lshl_b32 s4, s43, 16
; VI-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_mov_b32_e32 v5, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v11, v13, vcc
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    s_lshl_b32 s4, s45, 16
; VI-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[16:17], 16, v[15:16]
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_mov_b32_e32 v7, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v13, v15, vcc
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    s_lshl_b32 s4, s47, 16
; VI-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_and_b32 s4, s49, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_mov_b32_e32 v9, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v15, v17, vcc
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    s_lshl_b32 s4, s49, 16
; VI-NEXT:    v_cndmask_b32_e32 v13, v15, v17, vcc
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_and_b32 s4, s51, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_mov_b32_e32 v11, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v17, v19, vcc
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    s_lshl_b32 s4, s51, 16
; VI-NEXT:    v_cndmask_b32_e32 v15, v17, v19, vcc
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v13
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    s_and_b32 s4, s53, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_mov_b32_e32 v13, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v19, v21, vcc
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    s_lshl_b32 s4, s53, 16
; VI-NEXT:    v_cndmask_b32_e32 v17, v19, v21, vcc
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    s_and_b32 s4, s55, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_mov_b32_e32 v15, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v21, v23, vcc
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    s_lshl_b32 s4, s55, 16
; VI-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    s_and_b32 s4, s57, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_mov_b32_e32 v17, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v23, v25, vcc
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    s_lshl_b32 s4, s57, 16
; VI-NEXT:    v_cndmask_b32_e32 v21, v23, v25, vcc
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    s_and_b32 s4, s59, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_mov_b32_e32 v19, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v25, v27, vcc
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    s_lshl_b32 s4, s59, 16
; VI-NEXT:    v_cndmask_b32_e32 v23, v25, v27, vcc
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v21
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    s_and_b32 s4, s61, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_mov_b32_e32 v21, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v27, v29, vcc
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    s_lshl_b32 s4, s61, 16
; VI-NEXT:    v_cndmask_b32_e32 v25, v27, v29, vcc
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    s_and_b32 s4, s63, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_mov_b32_e32 v23, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v29, v31, vcc
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    s_lshl_b32 s4, s63, 16
; VI-NEXT:    v_cndmask_b32_e32 v27, v29, v31, vcc
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    s_and_b32 s4, s65, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v25, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v31, v33, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    v_mov_b32_e32 v27, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    s_lshl_b32 s4, s65, 16
; VI-NEXT:    v_cndmask_b32_e32 v29, v31, v33, vcc
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    s_and_b32 s4, s67, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_mov_b32_e32 v29, v33
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    s_lshl_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v32, v31, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_add_u32_e32 v32, vcc, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x7fff, v32
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v32, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; VI-NEXT:    v_lshrrev_b64 v[31:32], 16, v[31:32]
; VI-NEXT:    s_branch .LBB63_5
; VI-NEXT:  .LBB63_3:
; VI-NEXT:    s_branch .LBB63_2
; VI-NEXT:  .LBB63_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB63_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v35, 15
; VI-NEXT:    v_readlane_b32 s66, v35, 14
; VI-NEXT:    v_readlane_b32 s65, v35, 13
; VI-NEXT:    v_readlane_b32 s64, v35, 12
; VI-NEXT:    v_readlane_b32 s55, v35, 11
; VI-NEXT:    v_readlane_b32 s54, v35, 10
; VI-NEXT:    v_readlane_b32 s53, v35, 9
; VI-NEXT:    v_readlane_b32 s52, v35, 8
; VI-NEXT:    v_readlane_b32 s51, v35, 7
; VI-NEXT:    v_readlane_b32 s50, v35, 6
; VI-NEXT:    v_readlane_b32 s49, v35, 5
; VI-NEXT:    v_readlane_b32 s48, v35, 4
; VI-NEXT:    v_readlane_b32 s39, v35, 3
; VI-NEXT:    v_readlane_b32 s38, v35, 2
; VI-NEXT:    v_readlane_b32 s37, v35, 1
; VI-NEXT:    v_readlane_b32 s36, v35, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v16i64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v36, s36, 0
; GFX9-NEXT:    v_writelane_b32 v36, s37, 1
; GFX9-NEXT:    v_writelane_b32 v36, s38, 2
; GFX9-NEXT:    v_writelane_b32 v36, s39, 3
; GFX9-NEXT:    v_writelane_b32 v36, s48, 4
; GFX9-NEXT:    v_writelane_b32 v36, s49, 5
; GFX9-NEXT:    v_writelane_b32 v36, s50, 6
; GFX9-NEXT:    v_writelane_b32 v36, s51, 7
; GFX9-NEXT:    v_writelane_b32 v36, s52, 8
; GFX9-NEXT:    v_writelane_b32 v36, s53, 9
; GFX9-NEXT:    v_writelane_b32 v36, s54, 10
; GFX9-NEXT:    v_writelane_b32 v36, s55, 11
; GFX9-NEXT:    v_writelane_b32 v36, s64, 12
; GFX9-NEXT:    v_writelane_b32 v36, s65, 13
; GFX9-NEXT:    v_writelane_b32 v36, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v36, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB63_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB63_4
; GFX9-NEXT:  .LBB63_2: ; %cmp.true
; GFX9-NEXT:    s_and_b32 s4, s51, 0xffff0000
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x40c00000
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s51, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    s_and_b32 s4, s50, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_add_f32_e32 v2, s4, v16
; GFX9-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v3, v3, v2
; GFX9-NEXT:    v_add_u32_e32 v3, 0x7fff, v3
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    s_lshl_b32 s4, s50, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT:    v_add_f32_e32 v3, s4, v16
; GFX9-NEXT:    v_mov_b32_e32 v32, 0xffff
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    s_and_b32 s4, s49, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_lshl_or_b32 v15, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_and_b32_sdwa v3, v32, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_lshl_or_b32 v14, v2, 16, v3
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s49, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s48, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s48, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s47, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s47, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s46, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s46, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s45, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s45, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s44, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s44, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s43, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s43, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s42, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s42, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s41, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s41, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s40, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s40, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s39, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s39, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s38, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s38, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v17, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s37, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s37, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v17, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v1
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v17, v18, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s36, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v0
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s36, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v17, v18, vcc
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v17, v32, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s67, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s67, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s66, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v31, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s66, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s65, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v30, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s65, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s64, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v29, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s64, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s63, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v28, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s63, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s62, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v27, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s62, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s61, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v26, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s61, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s60, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s60, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s59, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v24, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s59, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s58, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v23, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s58, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s57, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v22, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s57, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s56, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v21, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s56, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s55, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v20, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s55, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v33, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s54, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v19, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s54, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v33, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v18
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s53, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v17
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s53, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v33, v32, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s52, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v33
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    s_lshl_b32 s4, s52, 16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v16, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v16, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v34, v35, vcc
; GFX9-NEXT:    v_and_b32_sdwa v16, v32, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v16, v32, 16, v16
; GFX9-NEXT:    s_branch .LBB63_5
; GFX9-NEXT:  .LBB63_3:
; GFX9-NEXT:    s_branch .LBB63_2
; GFX9-NEXT:  .LBB63_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB63_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v36, 15
; GFX9-NEXT:    v_readlane_b32 s66, v36, 14
; GFX9-NEXT:    v_readlane_b32 s65, v36, 13
; GFX9-NEXT:    v_readlane_b32 s64, v36, 12
; GFX9-NEXT:    v_readlane_b32 s55, v36, 11
; GFX9-NEXT:    v_readlane_b32 s54, v36, 10
; GFX9-NEXT:    v_readlane_b32 s53, v36, 9
; GFX9-NEXT:    v_readlane_b32 s52, v36, 8
; GFX9-NEXT:    v_readlane_b32 s51, v36, 7
; GFX9-NEXT:    v_readlane_b32 s50, v36, 6
; GFX9-NEXT:    v_readlane_b32 s49, v36, 5
; GFX9-NEXT:    v_readlane_b32 s48, v36, 4
; GFX9-NEXT:    v_readlane_b32 s39, v36, 3
; GFX9-NEXT:    v_readlane_b32 s38, v36, 2
; GFX9-NEXT:    v_readlane_b32 s37, v36, 1
; GFX9-NEXT:    v_readlane_b32 s36, v36, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v16i64_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s36, s0
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-TRUE16-NEXT:    s_mov_b32 s48, s24
; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
; GFX11-TRUE16-NEXT:    s_mov_b32 s39, s3
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-TRUE16-NEXT:    s_mov_b32 s49, s25
; GFX11-TRUE16-NEXT:    s_mov_b32 s38, s2
; GFX11-TRUE16-NEXT:    s_mov_b32 s37, s1
; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-TRUE16-NEXT:    s_mov_b32 s50, s26
; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-TRUE16-NEXT:    s_mov_b32 s51, s27
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-TRUE16-NEXT:    s_mov_b32 s52, s28
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-TRUE16-NEXT:    s_mov_b32 s53, s29
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB63_3
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB63_4
; GFX11-TRUE16-NEXT:  .LBB63_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s51, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s50, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s57, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v0.l
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v2.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v4, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v7, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v16, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v18, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v17 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v16, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v16 :: v_dual_add_nc_u32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v17, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v21, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s37, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v0.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v20 :: v_dual_add_nc_u32 v0, v16, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s36, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v22, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v16.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s67, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v24, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s66, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s65, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v20, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_nc_u32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s64, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s63, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s62, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s61, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v16.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s60, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s58, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v33, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s57, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v33 :: v_dual_add_nc_u32 v19, v19, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v32
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v33, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v20, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s55, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v20, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v35, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s52, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v19, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v33, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v19, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, v35, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v34, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v17.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v36, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v16, v34
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v36, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v35 :: v_dual_add_nc_u32 v35, v37, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v33
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v34, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v37, v38, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v17.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v35, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v32
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v16.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v33.l
; GFX11-TRUE16-NEXT:    s_branch .LBB63_5
; GFX11-TRUE16-NEXT:  .LBB63_3:
; GFX11-TRUE16-NEXT:    s_branch .LBB63_2
; GFX11-TRUE16-NEXT:  .LBB63_4:
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-TRUE16-NEXT:  .LBB63_5: ; %end
; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v16i64_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s36, s0
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-FAKE16-NEXT:    s_mov_b32 s47, s23
; GFX11-FAKE16-NEXT:    s_mov_b32 s46, s22
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-FAKE16-NEXT:    s_mov_b32 s45, s21
; GFX11-FAKE16-NEXT:    s_mov_b32 s44, s20
; GFX11-FAKE16-NEXT:    s_mov_b32 s43, s19
; GFX11-FAKE16-NEXT:    s_mov_b32 s42, s18
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-FAKE16-NEXT:    s_mov_b32 s48, s24
; GFX11-FAKE16-NEXT:    s_mov_b32 s41, s17
; GFX11-FAKE16-NEXT:    s_mov_b32 s40, s16
; GFX11-FAKE16-NEXT:    s_mov_b32 s39, s3
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-FAKE16-NEXT:    s_mov_b32 s49, s25
; GFX11-FAKE16-NEXT:    s_mov_b32 s38, s2
; GFX11-FAKE16-NEXT:    s_mov_b32 s37, s1
; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-FAKE16-NEXT:    s_mov_b32 s50, s26
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-FAKE16-NEXT:    s_mov_b32 s51, s27
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-FAKE16-NEXT:    s_mov_b32 s52, s28
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-FAKE16-NEXT:    s_mov_b32 s53, s29
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB63_3
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB63_4
; GFX11-FAKE16-NEXT:  .LBB63_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s51, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s50, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s49, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s49, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s48, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v7 :: v_dual_add_nc_u32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s47, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s46, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s45, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s44, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s43, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s42, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s41, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s40, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v17 :: v_dual_add_nc_u32 v5, v7, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s39, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s38, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s37, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v4, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s67, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v2, v22 :: v_dual_add_nc_u32 v17, v17, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v22, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s66, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s65, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v22, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s64, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s63, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s62, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s61, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s60, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s58, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v32, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v22, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v19, v21 :: v_dual_and_b32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s57, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v18, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s57, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v22, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v23, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v33, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v35, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v21, v33, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v33, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v22, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s55, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v34, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v33, v34 :: v_dual_add_nc_u32 v33, v35, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v37, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v33, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s52, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v38, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v35, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, v48, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v39
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v48
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v39, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v36, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v37, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v17, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v34, 16, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v32, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v36, 16, v33
; GFX11-FAKE16-NEXT:    s_branch .LBB63_5
; GFX11-FAKE16-NEXT:  .LBB63_3:
; GFX11-FAKE16-NEXT:    s_branch .LBB63_2
; GFX11-FAKE16-NEXT:  .LBB63_4:
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-FAKE16-NEXT:  .LBB63_5: ; %end
; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define <64 x half> @bitcast_v16i64_to_v64f16(<16 x i64> %a, i32 %b) {
; SI-LABEL: bitcast_v16i64_to_v64f16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB64_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v54, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v41, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB64_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB64_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v54, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v41, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB64_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v41
; SI-NEXT:    v_or_b32_e32 v0, v0, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v44
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v41
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v61
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v47
; SI-NEXT:    v_or_b32_e32 v3, v3, v44
; SI-NEXT:    v_or_b32_e32 v5, v5, v41
; SI-NEXT:    v_or_b32_e32 v6, v6, v54
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v60
; SI-NEXT:    v_or_b32_e32 v8, v8, v52
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v57
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v45
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v42
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v55
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v53
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v54
; SI-NEXT:    v_or_b32_e32 v9, v9, v52
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16i64_to_v64f16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB64_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; VI-NEXT:  .LBB64_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16i64_to_v64f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB64_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v14
; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v12
; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v10
; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v8
; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v6
; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v4
; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v2
; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v0
; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v28
; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v26
; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v24
; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v22
; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v20
; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v18, vcc, 3, v18
; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v16
; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
; GFX9-NEXT:  .LBB64_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16i64_to_v64f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB64_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 3
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
; GFX11-NEXT:  .LBB64_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define inreg <64 x half> @bitcast_v16i64_to_v64f16_scalar(<16 x i64> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16i64_to_v64f16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_readfirstlane_b32 s44, v18
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_readfirstlane_b32 s7, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    v_readfirstlane_b32 s9, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_readfirstlane_b32 s10, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s15, v7
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s40, v4
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    v_readfirstlane_b32 s42, v2
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    s_cbranch_scc0 .LBB65_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB65_3
; SI-NEXT:  .LBB65_2: ; %cmp.true
; SI-NEXT:    s_add_u32 s4, s4, 3
; SI-NEXT:    s_addc_u32 s5, s5, 0
; SI-NEXT:    s_add_u32 s6, s6, 3
; SI-NEXT:    s_addc_u32 s7, s7, 0
; SI-NEXT:    s_add_u32 s8, s8, 3
; SI-NEXT:    s_addc_u32 s9, s9, 0
; SI-NEXT:    s_add_u32 s10, s10, 3
; SI-NEXT:    s_addc_u32 s11, s11, 0
; SI-NEXT:    s_add_u32 s12, s12, 3
; SI-NEXT:    s_addc_u32 s13, s13, 0
; SI-NEXT:    s_add_u32 s14, s14, 3
; SI-NEXT:    s_addc_u32 s15, s15, 0
; SI-NEXT:    s_add_u32 s40, s40, 3
; SI-NEXT:    s_addc_u32 s41, s41, 0
; SI-NEXT:    s_add_u32 s42, s42, 3
; SI-NEXT:    s_addc_u32 s43, s43, 0
; SI-NEXT:    s_add_u32 s44, s44, 3
; SI-NEXT:    s_addc_u32 s45, s45, 0
; SI-NEXT:    s_add_u32 s28, s28, 3
; SI-NEXT:    s_addc_u32 s29, s29, 0
; SI-NEXT:    s_add_u32 s26, s26, 3
; SI-NEXT:    s_addc_u32 s27, s27, 0
; SI-NEXT:    s_add_u32 s24, s24, 3
; SI-NEXT:    s_addc_u32 s25, s25, 0
; SI-NEXT:    s_add_u32 s22, s22, 3
; SI-NEXT:    s_addc_u32 s23, s23, 0
; SI-NEXT:    s_add_u32 s20, s20, 3
; SI-NEXT:    s_addc_u32 s21, s21, 0
; SI-NEXT:    s_add_u32 s18, s18, 3
; SI-NEXT:    s_addc_u32 s19, s19, 0
; SI-NEXT:    s_add_u32 s16, s16, 3
; SI-NEXT:    s_addc_u32 s17, s17, 0
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:  .LBB65_3: ; %end
; SI-NEXT:    s_lshl_b32 s47, s36, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s69, 16
; SI-NEXT:    s_or_b32 s17, s17, s47
; SI-NEXT:    s_lshl_b32 s47, s34, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s47
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s68, 16
; SI-NEXT:    s_or_b32 s19, s19, s47
; SI-NEXT:    s_lshl_b32 s47, s30, 16
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s47
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s67, 16
; SI-NEXT:    s_or_b32 s21, s21, s47
; SI-NEXT:    s_lshl_b32 s47, s94, 16
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s47
; SI-NEXT:    s_and_b32 s23, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s66, 16
; SI-NEXT:    s_or_b32 s23, s23, s47
; SI-NEXT:    s_lshl_b32 s47, s92, 16
; SI-NEXT:    s_and_b32 s24, s24, 0xffff
; SI-NEXT:    s_or_b32 s24, s24, s47
; SI-NEXT:    s_and_b32 s25, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s65, 16
; SI-NEXT:    s_or_b32 s25, s25, s47
; SI-NEXT:    s_and_b32 s26, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s90, 16
; SI-NEXT:    s_or_b32 s26, s26, s47
; SI-NEXT:    s_and_b32 s27, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s64, 16
; SI-NEXT:    s_or_b32 s27, s27, s47
; SI-NEXT:    s_and_b32 s28, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s88, 16
; SI-NEXT:    s_or_b32 s28, s28, s47
; SI-NEXT:    s_and_b32 s29, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s55, 16
; SI-NEXT:    s_or_b32 s29, s29, s47
; SI-NEXT:    s_and_b32 s44, s44, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s78, 16
; SI-NEXT:    s_or_b32 s44, s44, s47
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s54, 16
; SI-NEXT:    s_or_b32 s45, s45, s47
; SI-NEXT:    s_and_b32 s42, s42, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s76, 16
; SI-NEXT:    s_or_b32 s42, s42, s47
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s53, 16
; SI-NEXT:    s_or_b32 s43, s43, s47
; SI-NEXT:    s_and_b32 s40, s40, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s74, 16
; SI-NEXT:    s_or_b32 s40, s40, s47
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s52, 16
; SI-NEXT:    s_or_b32 s41, s41, s47
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s72, 16
; SI-NEXT:    s_or_b32 s14, s14, s47
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s51, 16
; SI-NEXT:    s_or_b32 s15, s15, s47
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s62, 16
; SI-NEXT:    s_or_b32 s12, s12, s47
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s50, 16
; SI-NEXT:    s_or_b32 s13, s13, s47
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s58, 16
; SI-NEXT:    s_or_b32 s10, s10, s47
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s49, 16
; SI-NEXT:    s_or_b32 s11, s11, s47
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s60, 16
; SI-NEXT:    s_or_b32 s8, s8, s47
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s48, 16
; SI-NEXT:    s_or_b32 s9, s9, s47
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s56, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_or_b32 s6, s6, s47
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s39, 16
; SI-NEXT:    s_or_b32 s4, s4, s46
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s38, 16
; SI-NEXT:    s_or_b32 s7, s7, s47
; SI-NEXT:    s_or_b32 s5, s5, s46
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB65_4:
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    s_branch .LBB65_2
;
; VI-LABEL: bitcast_v16i64_to_v64f16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB65_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB65_3
; VI-NEXT:  .LBB65_2: ; %cmp.true
; VI-NEXT:    s_add_u32 s47, s47, 3
; VI-NEXT:    s_addc_u32 s46, s46, 0
; VI-NEXT:    s_add_u32 s28, s28, 3
; VI-NEXT:    s_addc_u32 s29, s29, 0
; VI-NEXT:    s_add_u32 s26, s26, 3
; VI-NEXT:    s_addc_u32 s27, s27, 0
; VI-NEXT:    s_add_u32 s24, s24, 3
; VI-NEXT:    s_addc_u32 s25, s25, 0
; VI-NEXT:    s_add_u32 s22, s22, 3
; VI-NEXT:    s_addc_u32 s23, s23, 0
; VI-NEXT:    s_add_u32 s20, s20, 3
; VI-NEXT:    s_addc_u32 s21, s21, 0
; VI-NEXT:    s_add_u32 s18, s18, 3
; VI-NEXT:    s_addc_u32 s19, s19, 0
; VI-NEXT:    s_add_u32 s16, s16, 3
; VI-NEXT:    s_addc_u32 s17, s17, 0
; VI-NEXT:    s_add_u32 s7, s7, 3
; VI-NEXT:    s_addc_u32 s6, s6, 0
; VI-NEXT:    s_add_u32 s9, s9, 3
; VI-NEXT:    s_addc_u32 s8, s8, 0
; VI-NEXT:    s_add_u32 s11, s11, 3
; VI-NEXT:    s_addc_u32 s10, s10, 0
; VI-NEXT:    s_add_u32 s13, s13, 3
; VI-NEXT:    s_addc_u32 s12, s12, 0
; VI-NEXT:    s_add_u32 s15, s15, 3
; VI-NEXT:    s_addc_u32 s14, s14, 0
; VI-NEXT:    s_add_u32 s41, s41, 3
; VI-NEXT:    s_addc_u32 s40, s40, 0
; VI-NEXT:    s_add_u32 s43, s43, 3
; VI-NEXT:    s_addc_u32 s42, s42, 0
; VI-NEXT:    s_add_u32 s45, s45, 3
; VI-NEXT:    s_addc_u32 s44, s44, 0
; VI-NEXT:  .LBB65_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB65_4:
; VI-NEXT:    s_branch .LBB65_2
;
; GFX9-LABEL: bitcast_v16i64_to_v64f16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB65_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB65_3
; GFX9-NEXT:  .LBB65_2: ; %cmp.true
; GFX9-NEXT:    s_add_u32 s47, s47, 3
; GFX9-NEXT:    s_addc_u32 s46, s46, 0
; GFX9-NEXT:    s_add_u32 s28, s28, 3
; GFX9-NEXT:    s_addc_u32 s29, s29, 0
; GFX9-NEXT:    s_add_u32 s26, s26, 3
; GFX9-NEXT:    s_addc_u32 s27, s27, 0
; GFX9-NEXT:    s_add_u32 s24, s24, 3
; GFX9-NEXT:    s_addc_u32 s25, s25, 0
; GFX9-NEXT:    s_add_u32 s22, s22, 3
; GFX9-NEXT:    s_addc_u32 s23, s23, 0
; GFX9-NEXT:    s_add_u32 s20, s20, 3
; GFX9-NEXT:    s_addc_u32 s21, s21, 0
; GFX9-NEXT:    s_add_u32 s18, s18, 3
; GFX9-NEXT:    s_addc_u32 s19, s19, 0
; GFX9-NEXT:    s_add_u32 s16, s16, 3
; GFX9-NEXT:    s_addc_u32 s17, s17, 0
; GFX9-NEXT:    s_add_u32 s7, s7, 3
; GFX9-NEXT:    s_addc_u32 s6, s6, 0
; GFX9-NEXT:    s_add_u32 s9, s9, 3
; GFX9-NEXT:    s_addc_u32 s8, s8, 0
; GFX9-NEXT:    s_add_u32 s11, s11, 3
; GFX9-NEXT:    s_addc_u32 s10, s10, 0
; GFX9-NEXT:    s_add_u32 s13, s13, 3
; GFX9-NEXT:    s_addc_u32 s12, s12, 0
; GFX9-NEXT:    s_add_u32 s15, s15, 3
; GFX9-NEXT:    s_addc_u32 s14, s14, 0
; GFX9-NEXT:    s_add_u32 s41, s41, 3
; GFX9-NEXT:    s_addc_u32 s40, s40, 0
; GFX9-NEXT:    s_add_u32 s43, s43, 3
; GFX9-NEXT:    s_addc_u32 s42, s42, 0
; GFX9-NEXT:    s_add_u32 s45, s45, 3
; GFX9-NEXT:    s_addc_u32 s44, s44, 0
; GFX9-NEXT:  .LBB65_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB65_4:
; GFX9-NEXT:    s_branch .LBB65_2
;
; GFX11-LABEL: bitcast_v16i64_to_v64f16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB65_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB65_3
; GFX11-NEXT:  .LBB65_2: ; %cmp.true
; GFX11-NEXT:    s_add_u32 s26, s26, 3
; GFX11-NEXT:    s_addc_u32 s27, s27, 0
; GFX11-NEXT:    s_add_u32 s24, s24, 3
; GFX11-NEXT:    s_addc_u32 s25, s25, 0
; GFX11-NEXT:    s_add_u32 s22, s22, 3
; GFX11-NEXT:    s_addc_u32 s23, s23, 0
; GFX11-NEXT:    s_add_u32 s20, s20, 3
; GFX11-NEXT:    s_addc_u32 s21, s21, 0
; GFX11-NEXT:    s_add_u32 s18, s18, 3
; GFX11-NEXT:    s_addc_u32 s19, s19, 0
; GFX11-NEXT:    s_add_u32 s16, s16, 3
; GFX11-NEXT:    s_addc_u32 s17, s17, 0
; GFX11-NEXT:    s_add_u32 s2, s2, 3
; GFX11-NEXT:    s_addc_u32 s3, s3, 0
; GFX11-NEXT:    s_add_u32 s0, s0, 3
; GFX11-NEXT:    s_addc_u32 s1, s1, 0
; GFX11-NEXT:    s_add_u32 s5, s5, 3
; GFX11-NEXT:    s_addc_u32 s4, s4, 0
; GFX11-NEXT:    s_add_u32 s7, s7, 3
; GFX11-NEXT:    s_addc_u32 s6, s6, 0
; GFX11-NEXT:    s_add_u32 s9, s9, 3
; GFX11-NEXT:    s_addc_u32 s8, s8, 0
; GFX11-NEXT:    s_add_u32 s11, s11, 3
; GFX11-NEXT:    s_addc_u32 s10, s10, 0
; GFX11-NEXT:    s_add_u32 s13, s13, 3
; GFX11-NEXT:    s_addc_u32 s12, s12, 0
; GFX11-NEXT:    s_add_u32 s15, s15, 3
; GFX11-NEXT:    s_addc_u32 s14, s14, 0
; GFX11-NEXT:    s_add_u32 s41, s41, 3
; GFX11-NEXT:    s_addc_u32 s40, s40, 0
; GFX11-NEXT:    s_add_u32 s28, s28, 3
; GFX11-NEXT:    s_addc_u32 s29, s29, 0
; GFX11-NEXT:  .LBB65_3: ; %end
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB65_4:
; GFX11-NEXT:    s_branch .LBB65_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define <16 x i64> @bitcast_v64f16_to_v16i64(<64 x half> %a, i32 %b) {
; SI-LABEL: bitcast_v64f16_to_v16i64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v30
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v29
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v27
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v25
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v24
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v23
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v22
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v21
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v20
; SI-NEXT:    v_mov_b32_e32 v52, v18
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
; SI-NEXT:    v_mov_b32_e32 v53, v17
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v52
; SI-NEXT:    v_mov_b32_e32 v54, v16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v53
; SI-NEXT:    v_mov_b32_e32 v55, v15
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v54
; SI-NEXT:    v_mov_b32_e32 v40, v14
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v55
; SI-NEXT:    v_mov_b32_e32 v41, v13
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v40
; SI-NEXT:    v_mov_b32_e32 v42, v12
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v28
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v42
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v61
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v62
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB66_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v63
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v61
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v32
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v36
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v49
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v33
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v48
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v35
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v39
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v38
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v37
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v42
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v41
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v40
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v55
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v53
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v52
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB66_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB66_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, v63
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v50
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v62
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v61
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_or_b32_e32 v1, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v59
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v58
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v57
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v56
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v49
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v47
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v46
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v33
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v45
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v44
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v48
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v43
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v42
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v35
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v41
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v40
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v55
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v54
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v34
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v53
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v19, v52
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v38
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v37
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v51
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:  .LBB66_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v16i64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB66_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x200
; VI-NEXT:    v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v15, 0x200, v15
; VI-NEXT:    v_or_b32_e32 v15, v15, v33
; VI-NEXT:    v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v14, 0x200, v14
; VI-NEXT:    v_or_b32_e32 v14, v14, v33
; VI-NEXT:    v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v13, 0x200, v13
; VI-NEXT:    v_or_b32_e32 v13, v13, v33
; VI-NEXT:    v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v12, 0x200, v12
; VI-NEXT:    v_or_b32_e32 v12, v12, v33
; VI-NEXT:    v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v11, 0x200, v11
; VI-NEXT:    v_or_b32_e32 v11, v11, v33
; VI-NEXT:    v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v10, 0x200, v10
; VI-NEXT:    v_or_b32_e32 v10, v10, v33
; VI-NEXT:    v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v9, 0x200, v9
; VI-NEXT:    v_or_b32_e32 v9, v9, v33
; VI-NEXT:    v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v8, 0x200, v8
; VI-NEXT:    v_or_b32_e32 v8, v8, v33
; VI-NEXT:    v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v7, 0x200, v7
; VI-NEXT:    v_or_b32_e32 v7, v7, v33
; VI-NEXT:    v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v6, 0x200, v6
; VI-NEXT:    v_or_b32_e32 v6, v6, v33
; VI-NEXT:    v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v5, 0x200, v5
; VI-NEXT:    v_or_b32_e32 v5, v5, v33
; VI-NEXT:    v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v4, 0x200, v4
; VI-NEXT:    v_or_b32_e32 v4, v4, v33
; VI-NEXT:    v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, 0x200, v3
; VI-NEXT:    v_or_b32_e32 v3, v3, v33
; VI-NEXT:    v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v2, 0x200, v2
; VI-NEXT:    v_or_b32_e32 v2, v2, v33
; VI-NEXT:    v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, 0x200, v1
; VI-NEXT:    v_or_b32_e32 v1, v1, v33
; VI-NEXT:    v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, 0x200, v0
; VI-NEXT:    v_or_b32_e32 v0, v0, v33
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v31, 0x200, v31
; VI-NEXT:    v_or_b32_e32 v31, v31, v33
; VI-NEXT:    v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v30, 0x200, v30
; VI-NEXT:    v_or_b32_e32 v30, v30, v33
; VI-NEXT:    v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v29, 0x200, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v28, 0x200, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v33
; VI-NEXT:    v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v27, 0x200, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v33
; VI-NEXT:    v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v26, 0x200, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v33
; VI-NEXT:    v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v25, 0x200, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v33
; VI-NEXT:    v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v24, 0x200, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v33
; VI-NEXT:    v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v23, 0x200, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v33
; VI-NEXT:    v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v22, 0x200, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v33
; VI-NEXT:    v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v21, 0x200, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v33
; VI-NEXT:    v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v20, 0x200, v20
; VI-NEXT:    v_or_b32_e32 v20, v20, v33
; VI-NEXT:    v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v19, 0x200, v19
; VI-NEXT:    v_or_b32_e32 v19, v19, v33
; VI-NEXT:    v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, 0x200, v18
; VI-NEXT:    v_or_b32_e32 v18, v18, v33
; VI-NEXT:    v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, 0x200, v17
; VI-NEXT:    v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, 0x200, v16
; VI-NEXT:    v_or_b32_e32 v17, v17, v33
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB66_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v16i64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB66_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_movk_i32 s6, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB66_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v16i64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB66_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-NEXT:  .LBB66_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define inreg <16 x i64> @bitcast_v64f16_to_v16i64_scalar(<64 x half> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64f16_to_v16i64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v34, s30, 0
; SI-NEXT:    v_writelane_b32 v34, s31, 1
; SI-NEXT:    v_writelane_b32 v34, s34, 2
; SI-NEXT:    v_writelane_b32 v34, s35, 3
; SI-NEXT:    v_writelane_b32 v34, s36, 4
; SI-NEXT:    v_writelane_b32 v34, s37, 5
; SI-NEXT:    v_writelane_b32 v34, s38, 6
; SI-NEXT:    v_writelane_b32 v34, s39, 7
; SI-NEXT:    v_writelane_b32 v34, s48, 8
; SI-NEXT:    v_writelane_b32 v34, s49, 9
; SI-NEXT:    v_writelane_b32 v34, s50, 10
; SI-NEXT:    v_writelane_b32 v34, s51, 11
; SI-NEXT:    v_writelane_b32 v34, s52, 12
; SI-NEXT:    v_writelane_b32 v34, s53, 13
; SI-NEXT:    v_writelane_b32 v34, s54, 14
; SI-NEXT:    v_writelane_b32 v34, s55, 15
; SI-NEXT:    v_writelane_b32 v34, s64, 16
; SI-NEXT:    v_writelane_b32 v34, s65, 17
; SI-NEXT:    v_writelane_b32 v34, s66, 18
; SI-NEXT:    v_writelane_b32 v34, s67, 19
; SI-NEXT:    v_writelane_b32 v34, s68, 20
; SI-NEXT:    v_writelane_b32 v34, s69, 21
; SI-NEXT:    v_writelane_b32 v34, s70, 22
; SI-NEXT:    v_writelane_b32 v34, s71, 23
; SI-NEXT:    v_writelane_b32 v34, s80, 24
; SI-NEXT:    v_writelane_b32 v34, s81, 25
; SI-NEXT:    v_writelane_b32 v34, s82, 26
; SI-NEXT:    v_writelane_b32 v34, s83, 27
; SI-NEXT:    v_writelane_b32 v34, s84, 28
; SI-NEXT:    v_writelane_b32 v34, s85, 29
; SI-NEXT:    v_writelane_b32 v34, s86, 30
; SI-NEXT:    v_writelane_b32 v34, s87, 31
; SI-NEXT:    v_writelane_b32 v34, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_writelane_b32 v34, s97, 33
; SI-NEXT:    s_lshr_b32 vcc_lo, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s8, v16
; SI-NEXT:    ; implicit-def: $vgpr35 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v34, s98, 34
; SI-NEXT:    s_lshr_b32 vcc_hi, s8, 16
; SI-NEXT:    v_readfirstlane_b32 s10, v15
; SI-NEXT:    v_readfirstlane_b32 s12, v14
; SI-NEXT:    v_readfirstlane_b32 s14, v13
; SI-NEXT:    v_readfirstlane_b32 s72, v12
; SI-NEXT:    v_readfirstlane_b32 s74, v11
; SI-NEXT:    v_readfirstlane_b32 s76, v10
; SI-NEXT:    v_readfirstlane_b32 s78, v9
; SI-NEXT:    v_readfirstlane_b32 s88, v8
; SI-NEXT:    v_readfirstlane_b32 s90, v7
; SI-NEXT:    v_readfirstlane_b32 s92, v6
; SI-NEXT:    v_readfirstlane_b32 s94, v5
; SI-NEXT:    v_readfirstlane_b32 s30, v4
; SI-NEXT:    v_readfirstlane_b32 s34, v3
; SI-NEXT:    v_readfirstlane_b32 s68, v2
; SI-NEXT:    v_readfirstlane_b32 s80, v1
; SI-NEXT:    v_readfirstlane_b32 s83, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v35, vcc_lo, 0
; SI-NEXT:    v_writelane_b32 v34, s99, 35
; SI-NEXT:    s_lshr_b32 s69, s29, 16
; SI-NEXT:    s_lshr_b32 s71, s28, 16
; SI-NEXT:    s_lshr_b32 s82, s27, 16
; SI-NEXT:    s_lshr_b32 s85, s26, 16
; SI-NEXT:    s_lshr_b32 s86, s25, 16
; SI-NEXT:    s_lshr_b32 s87, s24, 16
; SI-NEXT:    s_lshr_b32 s96, s23, 16
; SI-NEXT:    s_lshr_b32 s97, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s7, s19, 16
; SI-NEXT:    s_lshr_b32 s9, s18, 16
; SI-NEXT:    s_lshr_b32 s11, s17, 16
; SI-NEXT:    s_lshr_b32 s13, s16, 16
; SI-NEXT:    s_lshr_b32 s65, s10, 16
; SI-NEXT:    s_lshr_b32 s64, s12, 16
; SI-NEXT:    s_lshr_b32 s15, s14, 16
; SI-NEXT:    s_lshr_b32 s73, s72, 16
; SI-NEXT:    s_lshr_b32 s75, s74, 16
; SI-NEXT:    s_lshr_b32 s77, s76, 16
; SI-NEXT:    s_lshr_b32 s79, s78, 16
; SI-NEXT:    s_lshr_b32 s89, s88, 16
; SI-NEXT:    s_lshr_b32 s91, s90, 16
; SI-NEXT:    s_lshr_b32 s93, s92, 16
; SI-NEXT:    s_lshr_b32 s95, s94, 16
; SI-NEXT:    s_lshr_b32 s31, s30, 16
; SI-NEXT:    s_lshr_b32 s35, s34, 16
; SI-NEXT:    s_lshr_b32 s70, s68, 16
; SI-NEXT:    s_lshr_b32 s81, s80, 16
; SI-NEXT:    s_lshr_b32 s84, s83, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v35, vcc_hi, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v35, s65, 2
; SI-NEXT:    v_writelane_b32 v35, s64, 3
; SI-NEXT:    s_cbranch_scc0 .LBB67_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s13, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s9, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s7, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s11, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s97, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s86, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s34, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s30, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s31, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s92, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s90, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s91, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s88, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s89, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s78, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s79, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s76, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s77, 16
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s74, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s75, 16
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s72, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s73, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s15, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s64, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s65, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB67_4
; SI-NEXT:  .LBB67_2: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s13
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s16
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s11
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s17
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s18
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s9
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v1, v3, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s7
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s19
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s99
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s20
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s98
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s21
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s97
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v7, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s22
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s96
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s23
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s87
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s24
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_or_b32_e32 v7, v9, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s86
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v10, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s25
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s85
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s26
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s82
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s27
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_or_b32_e32 v10, v12, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s71
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v13, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s28
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s69
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s29
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s84
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s83
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_or_b32_e32 v13, v15, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s81
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s80
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s70
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s68
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s35
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s34
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_or_b32_e32 v16, v18, v16
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s31
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s30
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s95
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s94
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s93
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s92
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s91
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s90
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s89
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s88
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s79
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s78
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s77
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s76
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s75
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s74
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s73
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s72
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s15
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_readlane_b32 s4, v35, 3
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s14
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_readlane_b32 s4, v35, 2
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s12
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s10
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_readlane_b32 s4, v35, 1
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s4
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_readlane_b32 s4, v35, 0
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s8
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v33, s6
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:    s_branch .LBB67_5
; SI-NEXT:  .LBB67_3:
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB67_2
; SI-NEXT:  .LBB67_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB67_5: ; %end
; SI-NEXT:    v_readlane_b32 s99, v34, 35
; SI-NEXT:    v_readlane_b32 s98, v34, 34
; SI-NEXT:    v_readlane_b32 s97, v34, 33
; SI-NEXT:    v_readlane_b32 s96, v34, 32
; SI-NEXT:    v_readlane_b32 s87, v34, 31
; SI-NEXT:    v_readlane_b32 s86, v34, 30
; SI-NEXT:    v_readlane_b32 s85, v34, 29
; SI-NEXT:    v_readlane_b32 s84, v34, 28
; SI-NEXT:    v_readlane_b32 s83, v34, 27
; SI-NEXT:    v_readlane_b32 s82, v34, 26
; SI-NEXT:    v_readlane_b32 s81, v34, 25
; SI-NEXT:    v_readlane_b32 s80, v34, 24
; SI-NEXT:    v_readlane_b32 s71, v34, 23
; SI-NEXT:    v_readlane_b32 s70, v34, 22
; SI-NEXT:    v_readlane_b32 s69, v34, 21
; SI-NEXT:    v_readlane_b32 s68, v34, 20
; SI-NEXT:    v_readlane_b32 s67, v34, 19
; SI-NEXT:    v_readlane_b32 s66, v34, 18
; SI-NEXT:    v_readlane_b32 s65, v34, 17
; SI-NEXT:    v_readlane_b32 s64, v34, 16
; SI-NEXT:    v_readlane_b32 s55, v34, 15
; SI-NEXT:    v_readlane_b32 s54, v34, 14
; SI-NEXT:    v_readlane_b32 s53, v34, 13
; SI-NEXT:    v_readlane_b32 s52, v34, 12
; SI-NEXT:    v_readlane_b32 s51, v34, 11
; SI-NEXT:    v_readlane_b32 s50, v34, 10
; SI-NEXT:    v_readlane_b32 s49, v34, 9
; SI-NEXT:    v_readlane_b32 s48, v34, 8
; SI-NEXT:    v_readlane_b32 s39, v34, 7
; SI-NEXT:    v_readlane_b32 s38, v34, 6
; SI-NEXT:    v_readlane_b32 s37, v34, 5
; SI-NEXT:    v_readlane_b32 s36, v34, 4
; SI-NEXT:    v_readlane_b32 s35, v34, 3
; SI-NEXT:    v_readlane_b32 s34, v34, 2
; SI-NEXT:    v_readlane_b32 s31, v34, 1
; SI-NEXT:    v_readlane_b32 s30, v34, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v16i64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v33, s36, 0
; VI-NEXT:    v_writelane_b32 v33, s37, 1
; VI-NEXT:    v_writelane_b32 v33, s38, 2
; VI-NEXT:    v_writelane_b32 v33, s39, 3
; VI-NEXT:    v_writelane_b32 v33, s48, 4
; VI-NEXT:    v_writelane_b32 v33, s49, 5
; VI-NEXT:    v_writelane_b32 v33, s50, 6
; VI-NEXT:    v_writelane_b32 v33, s51, 7
; VI-NEXT:    v_writelane_b32 v33, s52, 8
; VI-NEXT:    v_writelane_b32 v33, s53, 9
; VI-NEXT:    v_writelane_b32 v33, s54, 10
; VI-NEXT:    v_writelane_b32 v33, s55, 11
; VI-NEXT:    v_writelane_b32 v33, s64, 12
; VI-NEXT:    v_writelane_b32 v33, s65, 13
; VI-NEXT:    v_writelane_b32 v33, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v33, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB67_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB67_4
; VI-NEXT:  .LBB67_2: ; %cmp.true
; VI-NEXT:    s_lshr_b32 s4, s51, 16
; VI-NEXT:    v_mov_b32_e32 v16, 0x200
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    s_lshr_b32 s4, s50, 16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s51, v16
; VI-NEXT:    v_mov_b32_e32 v2, s4
; VI-NEXT:    s_lshr_b32 s4, s49, 16
; VI-NEXT:    v_or_b32_e32 v15, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s49, v16
; VI-NEXT:    s_lshr_b32 s4, s48, 16
; VI-NEXT:    v_or_b32_e32 v13, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s48, v16
; VI-NEXT:    s_lshr_b32 s4, s47, 16
; VI-NEXT:    v_or_b32_e32 v12, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s47, v16
; VI-NEXT:    s_lshr_b32 s4, s46, 16
; VI-NEXT:    v_or_b32_e32 v11, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s46, v16
; VI-NEXT:    s_lshr_b32 s4, s45, 16
; VI-NEXT:    v_or_b32_e32 v10, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s45, v16
; VI-NEXT:    s_lshr_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v9, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s44, v16
; VI-NEXT:    s_lshr_b32 s4, s43, 16
; VI-NEXT:    v_or_b32_e32 v8, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s43, v16
; VI-NEXT:    s_lshr_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v7, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s42, v16
; VI-NEXT:    s_lshr_b32 s4, s41, 16
; VI-NEXT:    v_or_b32_e32 v6, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s41, v16
; VI-NEXT:    s_lshr_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v5, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s40, v16
; VI-NEXT:    s_lshr_b32 s4, s39, 16
; VI-NEXT:    v_or_b32_e32 v4, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v2, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, s50, v16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s39, v16
; VI-NEXT:    s_lshr_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v14, v3, v2
; VI-NEXT:    v_or_b32_e32 v3, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s38, v16
; VI-NEXT:    s_lshr_b32 s4, s37, 16
; VI-NEXT:    v_or_b32_e32 v2, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s37, v16
; VI-NEXT:    s_lshr_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v1, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, s36, v16
; VI-NEXT:    s_lshr_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v0, v17, v0
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s67, v16
; VI-NEXT:    s_lshr_b32 s4, s66, 16
; VI-NEXT:    v_or_b32_e32 v31, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s66, v16
; VI-NEXT:    s_lshr_b32 s4, s65, 16
; VI-NEXT:    v_or_b32_e32 v30, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s65, v16
; VI-NEXT:    s_lshr_b32 s4, s64, 16
; VI-NEXT:    v_or_b32_e32 v29, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s64, v16
; VI-NEXT:    s_lshr_b32 s4, s63, 16
; VI-NEXT:    v_or_b32_e32 v28, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s63, v16
; VI-NEXT:    s_lshr_b32 s4, s62, 16
; VI-NEXT:    v_or_b32_e32 v27, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s62, v16
; VI-NEXT:    s_lshr_b32 s4, s61, 16
; VI-NEXT:    v_or_b32_e32 v26, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s61, v16
; VI-NEXT:    s_lshr_b32 s4, s60, 16
; VI-NEXT:    v_or_b32_e32 v25, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s60, v16
; VI-NEXT:    s_lshr_b32 s4, s59, 16
; VI-NEXT:    v_or_b32_e32 v24, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s59, v16
; VI-NEXT:    s_lshr_b32 s4, s58, 16
; VI-NEXT:    v_or_b32_e32 v23, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s58, v16
; VI-NEXT:    s_lshr_b32 s4, s57, 16
; VI-NEXT:    v_or_b32_e32 v22, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s57, v16
; VI-NEXT:    s_lshr_b32 s4, s56, 16
; VI-NEXT:    v_or_b32_e32 v21, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s56, v16
; VI-NEXT:    s_lshr_b32 s4, s55, 16
; VI-NEXT:    v_or_b32_e32 v20, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s55, v16
; VI-NEXT:    s_lshr_b32 s4, s54, 16
; VI-NEXT:    v_or_b32_e32 v19, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s54, v16
; VI-NEXT:    s_lshr_b32 s4, s53, 16
; VI-NEXT:    v_or_b32_e32 v18, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v32, s53, v16
; VI-NEXT:    s_lshr_b32 s4, s52, 16
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_mov_b32_e32 v32, s4
; VI-NEXT:    v_add_f16_sdwa v32, v32, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, s52, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:    s_branch .LBB67_5
; VI-NEXT:  .LBB67_3:
; VI-NEXT:    s_branch .LBB67_2
; VI-NEXT:  .LBB67_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB67_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v33, 15
; VI-NEXT:    v_readlane_b32 s66, v33, 14
; VI-NEXT:    v_readlane_b32 s65, v33, 13
; VI-NEXT:    v_readlane_b32 s64, v33, 12
; VI-NEXT:    v_readlane_b32 s55, v33, 11
; VI-NEXT:    v_readlane_b32 s54, v33, 10
; VI-NEXT:    v_readlane_b32 s53, v33, 9
; VI-NEXT:    v_readlane_b32 s52, v33, 8
; VI-NEXT:    v_readlane_b32 s51, v33, 7
; VI-NEXT:    v_readlane_b32 s50, v33, 6
; VI-NEXT:    v_readlane_b32 s49, v33, 5
; VI-NEXT:    v_readlane_b32 s48, v33, 4
; VI-NEXT:    v_readlane_b32 s39, v33, 3
; VI-NEXT:    v_readlane_b32 s38, v33, 2
; VI-NEXT:    v_readlane_b32 s37, v33, 1
; VI-NEXT:    v_readlane_b32 s36, v33, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v16i64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB67_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB67_4
; GFX9-NEXT:  .LBB67_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, s51, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, s50, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, s49, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, s48, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, s47, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, s46, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, s45, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, s44, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, s43, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, s42, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, s41, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, s40, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, s39, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, s38, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, s37, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, s36, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v31, s67, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, s66, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, s65, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, s64, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, s63, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, s62, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, s61, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, s60, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, s59, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, s58, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, s57, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, s56, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, s55, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, s54, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, s53, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, s52, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB67_5
; GFX9-NEXT:  .LBB67_3:
; GFX9-NEXT:    s_branch .LBB67_2
; GFX9-NEXT:  .LBB67_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB67_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v16i64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB67_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB67_4
; GFX11-NEXT:  .LBB67_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s51 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s50 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s49 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s48 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s47 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s46 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s45 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s44 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s43 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s42 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s41 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s40 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s39 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s38 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s37 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s36 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, s67 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, s66 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s65 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s64 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, s63 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, s62 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, s61 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s60 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s59 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, s58 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s57 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s56 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s55 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s54 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s53 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s52 op_sel_hi:[0,1]
; GFX11-NEXT:    s_branch .LBB67_5
; GFX11-NEXT:  .LBB67_3:
; GFX11-NEXT:    s_branch .LBB67_2
; GFX11-NEXT:  .LBB67_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB67_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define <64 x i16> @bitcast_v16i64_to_v64i16(<16 x i64> %a, i32 %b) {
; SI-LABEL: bitcast_v16i64_to_v64i16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB68_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v54, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v41, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB68_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB68_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v54, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v41, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v44, v3, v2, 16
; SI-NEXT:    v_alignbit_b32 v47, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v15
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB68_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v41
; SI-NEXT:    v_or_b32_e32 v0, v0, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v44
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v41
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v61
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v47
; SI-NEXT:    v_or_b32_e32 v3, v3, v44
; SI-NEXT:    v_or_b32_e32 v5, v5, v41
; SI-NEXT:    v_or_b32_e32 v6, v6, v54
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v60
; SI-NEXT:    v_or_b32_e32 v8, v8, v52
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v57
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v45
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v42
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v55
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v53
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v54
; SI-NEXT:    v_or_b32_e32 v9, v9, v52
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16i64_to_v64i16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB68_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v14
; VI-NEXT:    v_addc_u32_e32 v15, vcc, 0, v15, vcc
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    v_addc_u32_e32 v13, vcc, 0, v13, vcc
; VI-NEXT:    v_add_u32_e32 v10, vcc, 3, v10
; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v11, vcc
; VI-NEXT:    v_add_u32_e32 v8, vcc, 3, v8
; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
; VI-NEXT:    v_add_u32_e32 v6, vcc, 3, v6
; VI-NEXT:    v_addc_u32_e32 v7, vcc, 0, v7, vcc
; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v4
; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v5, vcc
; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v2
; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_addc_u32_e32 v31, vcc, 0, v31, vcc
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_addc_u32_e32 v29, vcc, 0, v29, vcc
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_addc_u32_e32 v27, vcc, 0, v27, vcc
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_addc_u32_e32 v25, vcc, 0, v25, vcc
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_addc_u32_e32 v23, vcc, 0, v23, vcc
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_addc_u32_e32 v21, vcc, 0, v21, vcc
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v18
; VI-NEXT:    v_addc_u32_e32 v19, vcc, 0, v19, vcc
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v16
; VI-NEXT:    v_addc_u32_e32 v17, vcc, 0, v17, vcc
; VI-NEXT:  .LBB68_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16i64_to_v64i16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB68_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_co_u32_e32 v14, vcc, 3, v14
; GFX9-NEXT:    v_addc_co_u32_e32 v15, vcc, 0, v15, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v12, vcc, 3, v12
; GFX9-NEXT:    v_addc_co_u32_e32 v13, vcc, 0, v13, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v10, vcc, 3, v10
; GFX9-NEXT:    v_addc_co_u32_e32 v11, vcc, 0, v11, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v8, vcc, 3, v8
; GFX9-NEXT:    v_addc_co_u32_e32 v9, vcc, 0, v9, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v6, vcc, 3, v6
; GFX9-NEXT:    v_addc_co_u32_e32 v7, vcc, 0, v7, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v4, vcc, 3, v4
; GFX9-NEXT:    v_addc_co_u32_e32 v5, vcc, 0, v5, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v2, vcc, 3, v2
; GFX9-NEXT:    v_addc_co_u32_e32 v3, vcc, 0, v3, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, 3, v0
; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, 0, v1, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v30, vcc, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_addc_co_u32_e32 v31, vcc, 0, v31, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v28, vcc, 3, v28
; GFX9-NEXT:    v_addc_co_u32_e32 v29, vcc, 0, v29, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v26, vcc, 3, v26
; GFX9-NEXT:    v_addc_co_u32_e32 v27, vcc, 0, v27, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v24, vcc, 3, v24
; GFX9-NEXT:    v_addc_co_u32_e32 v25, vcc, 0, v25, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v22, vcc, 3, v22
; GFX9-NEXT:    v_addc_co_u32_e32 v23, vcc, 0, v23, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v20, vcc, 3, v20
; GFX9-NEXT:    v_addc_co_u32_e32 v21, vcc, 0, v21, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v18, vcc, 3, v18
; GFX9-NEXT:    v_addc_co_u32_e32 v19, vcc, 0, v19, vcc
; GFX9-NEXT:    v_add_co_u32_e32 v16, vcc, 3, v16
; GFX9-NEXT:    v_addc_co_u32_e32 v17, vcc, 0, v17, vcc
; GFX9-NEXT:  .LBB68_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16i64_to_v64i16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB68_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_co_u32 v14, vcc_lo, v14, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v15, null, 0, v15, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v12, vcc_lo, v12, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v13, null, 0, v13, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v10, vcc_lo, v10, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v11, null, 0, v11, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v8, vcc_lo, v8, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v9, null, 0, v9, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v6, vcc_lo, v6, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v7, null, 0, v7, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v4, vcc_lo, v4, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v5, null, 0, v5, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, 0, v3, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, v1, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v30, vcc_lo, v30, 3
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v31, null, 0, v31, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v28, vcc_lo, v28, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v29, null, 0, v29, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v26, vcc_lo, v26, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v27, null, 0, v27, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v24, vcc_lo, v24, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v25, null, 0, v25, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v22, vcc_lo, v22, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v23, null, 0, v23, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v20, vcc_lo, v20, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v21, null, 0, v21, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v18, vcc_lo, v18, 3
; GFX11-NEXT:    v_add_co_ci_u32_e64 v19, null, 0, v19, vcc_lo
; GFX11-NEXT:    v_add_co_u32 v16, vcc_lo, v16, 3
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-NEXT:    v_add_co_ci_u32_e64 v17, null, 0, v17, vcc_lo
; GFX11-NEXT:  .LBB68_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define inreg <64 x i16> @bitcast_v16i64_to_v64i16_scalar(<16 x i64> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16i64_to_v64i16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_readfirstlane_b32 s44, v18
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_readfirstlane_b32 s7, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    v_readfirstlane_b32 s9, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_readfirstlane_b32 s10, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s15, v7
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s40, v4
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    v_readfirstlane_b32 s42, v2
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    s_cbranch_scc0 .LBB69_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB69_3
; SI-NEXT:  .LBB69_2: ; %cmp.true
; SI-NEXT:    s_add_u32 s4, s4, 3
; SI-NEXT:    s_addc_u32 s5, s5, 0
; SI-NEXT:    s_add_u32 s6, s6, 3
; SI-NEXT:    s_addc_u32 s7, s7, 0
; SI-NEXT:    s_add_u32 s8, s8, 3
; SI-NEXT:    s_addc_u32 s9, s9, 0
; SI-NEXT:    s_add_u32 s10, s10, 3
; SI-NEXT:    s_addc_u32 s11, s11, 0
; SI-NEXT:    s_add_u32 s12, s12, 3
; SI-NEXT:    s_addc_u32 s13, s13, 0
; SI-NEXT:    s_add_u32 s14, s14, 3
; SI-NEXT:    s_addc_u32 s15, s15, 0
; SI-NEXT:    s_add_u32 s40, s40, 3
; SI-NEXT:    s_addc_u32 s41, s41, 0
; SI-NEXT:    s_add_u32 s42, s42, 3
; SI-NEXT:    s_addc_u32 s43, s43, 0
; SI-NEXT:    s_add_u32 s44, s44, 3
; SI-NEXT:    s_addc_u32 s45, s45, 0
; SI-NEXT:    s_add_u32 s28, s28, 3
; SI-NEXT:    s_addc_u32 s29, s29, 0
; SI-NEXT:    s_add_u32 s26, s26, 3
; SI-NEXT:    s_addc_u32 s27, s27, 0
; SI-NEXT:    s_add_u32 s24, s24, 3
; SI-NEXT:    s_addc_u32 s25, s25, 0
; SI-NEXT:    s_add_u32 s22, s22, 3
; SI-NEXT:    s_addc_u32 s23, s23, 0
; SI-NEXT:    s_add_u32 s20, s20, 3
; SI-NEXT:    s_addc_u32 s21, s21, 0
; SI-NEXT:    s_add_u32 s18, s18, 3
; SI-NEXT:    s_addc_u32 s19, s19, 0
; SI-NEXT:    s_add_u32 s16, s16, 3
; SI-NEXT:    s_addc_u32 s17, s17, 0
; SI-NEXT:    s_lshr_b32 s38, s5, 16
; SI-NEXT:    s_lshr_b32 s39, s7, 16
; SI-NEXT:    s_lshr_b32 s48, s9, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s50, s13, 16
; SI-NEXT:    s_lshr_b32 s51, s15, 16
; SI-NEXT:    s_lshr_b32 s52, s41, 16
; SI-NEXT:    s_lshr_b32 s53, s43, 16
; SI-NEXT:    s_lshr_b32 s54, s45, 16
; SI-NEXT:    s_lshr_b32 s55, s29, 16
; SI-NEXT:    s_lshr_b32 s64, s27, 16
; SI-NEXT:    s_lshr_b32 s65, s25, 16
; SI-NEXT:    s_lshr_b32 s66, s23, 16
; SI-NEXT:    s_lshr_b32 s67, s21, 16
; SI-NEXT:    s_lshr_b32 s68, s19, 16
; SI-NEXT:    s_lshr_b32 s69, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:  .LBB69_3: ; %end
; SI-NEXT:    s_lshl_b32 s47, s36, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s47
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s69, 16
; SI-NEXT:    s_or_b32 s17, s17, s47
; SI-NEXT:    s_lshl_b32 s47, s34, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s47
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s68, 16
; SI-NEXT:    s_or_b32 s19, s19, s47
; SI-NEXT:    s_lshl_b32 s47, s30, 16
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s47
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s67, 16
; SI-NEXT:    s_or_b32 s21, s21, s47
; SI-NEXT:    s_lshl_b32 s47, s94, 16
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s47
; SI-NEXT:    s_and_b32 s23, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s66, 16
; SI-NEXT:    s_or_b32 s23, s23, s47
; SI-NEXT:    s_lshl_b32 s47, s92, 16
; SI-NEXT:    s_and_b32 s24, s24, 0xffff
; SI-NEXT:    s_or_b32 s24, s24, s47
; SI-NEXT:    s_and_b32 s25, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s65, 16
; SI-NEXT:    s_or_b32 s25, s25, s47
; SI-NEXT:    s_and_b32 s26, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s90, 16
; SI-NEXT:    s_or_b32 s26, s26, s47
; SI-NEXT:    s_and_b32 s27, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s64, 16
; SI-NEXT:    s_or_b32 s27, s27, s47
; SI-NEXT:    s_and_b32 s28, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s88, 16
; SI-NEXT:    s_or_b32 s28, s28, s47
; SI-NEXT:    s_and_b32 s29, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s55, 16
; SI-NEXT:    s_or_b32 s29, s29, s47
; SI-NEXT:    s_and_b32 s44, s44, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s78, 16
; SI-NEXT:    s_or_b32 s44, s44, s47
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s54, 16
; SI-NEXT:    s_or_b32 s45, s45, s47
; SI-NEXT:    s_and_b32 s42, s42, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s76, 16
; SI-NEXT:    s_or_b32 s42, s42, s47
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s53, 16
; SI-NEXT:    s_or_b32 s43, s43, s47
; SI-NEXT:    s_and_b32 s40, s40, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s74, 16
; SI-NEXT:    s_or_b32 s40, s40, s47
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s52, 16
; SI-NEXT:    s_or_b32 s41, s41, s47
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s72, 16
; SI-NEXT:    s_or_b32 s14, s14, s47
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s51, 16
; SI-NEXT:    s_or_b32 s15, s15, s47
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s62, 16
; SI-NEXT:    s_or_b32 s12, s12, s47
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s50, 16
; SI-NEXT:    s_or_b32 s13, s13, s47
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s58, 16
; SI-NEXT:    s_or_b32 s10, s10, s47
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s49, 16
; SI-NEXT:    s_or_b32 s11, s11, s47
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s60, 16
; SI-NEXT:    s_or_b32 s8, s8, s47
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s48, 16
; SI-NEXT:    s_or_b32 s9, s9, s47
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s56, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_or_b32 s6, s6, s47
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s39, 16
; SI-NEXT:    s_or_b32 s4, s4, s46
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s38, 16
; SI-NEXT:    s_or_b32 s7, s7, s47
; SI-NEXT:    s_or_b32 s5, s5, s46
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB69_4:
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    s_branch .LBB69_2
;
; VI-LABEL: bitcast_v16i64_to_v64i16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB69_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB69_3
; VI-NEXT:  .LBB69_2: ; %cmp.true
; VI-NEXT:    s_add_u32 s47, s47, 3
; VI-NEXT:    s_addc_u32 s46, s46, 0
; VI-NEXT:    s_add_u32 s28, s28, 3
; VI-NEXT:    s_addc_u32 s29, s29, 0
; VI-NEXT:    s_add_u32 s26, s26, 3
; VI-NEXT:    s_addc_u32 s27, s27, 0
; VI-NEXT:    s_add_u32 s24, s24, 3
; VI-NEXT:    s_addc_u32 s25, s25, 0
; VI-NEXT:    s_add_u32 s22, s22, 3
; VI-NEXT:    s_addc_u32 s23, s23, 0
; VI-NEXT:    s_add_u32 s20, s20, 3
; VI-NEXT:    s_addc_u32 s21, s21, 0
; VI-NEXT:    s_add_u32 s18, s18, 3
; VI-NEXT:    s_addc_u32 s19, s19, 0
; VI-NEXT:    s_add_u32 s16, s16, 3
; VI-NEXT:    s_addc_u32 s17, s17, 0
; VI-NEXT:    s_add_u32 s7, s7, 3
; VI-NEXT:    s_addc_u32 s6, s6, 0
; VI-NEXT:    s_add_u32 s9, s9, 3
; VI-NEXT:    s_addc_u32 s8, s8, 0
; VI-NEXT:    s_add_u32 s11, s11, 3
; VI-NEXT:    s_addc_u32 s10, s10, 0
; VI-NEXT:    s_add_u32 s13, s13, 3
; VI-NEXT:    s_addc_u32 s12, s12, 0
; VI-NEXT:    s_add_u32 s15, s15, 3
; VI-NEXT:    s_addc_u32 s14, s14, 0
; VI-NEXT:    s_add_u32 s41, s41, 3
; VI-NEXT:    s_addc_u32 s40, s40, 0
; VI-NEXT:    s_add_u32 s43, s43, 3
; VI-NEXT:    s_addc_u32 s42, s42, 0
; VI-NEXT:    s_add_u32 s45, s45, 3
; VI-NEXT:    s_addc_u32 s44, s44, 0
; VI-NEXT:  .LBB69_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB69_4:
; VI-NEXT:    s_branch .LBB69_2
;
; GFX9-LABEL: bitcast_v16i64_to_v64i16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    v_readfirstlane_b32 s46, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s47, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB69_4
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB69_3
; GFX9-NEXT:  .LBB69_2: ; %cmp.true
; GFX9-NEXT:    s_add_u32 s47, s47, 3
; GFX9-NEXT:    s_addc_u32 s46, s46, 0
; GFX9-NEXT:    s_add_u32 s28, s28, 3
; GFX9-NEXT:    s_addc_u32 s29, s29, 0
; GFX9-NEXT:    s_add_u32 s26, s26, 3
; GFX9-NEXT:    s_addc_u32 s27, s27, 0
; GFX9-NEXT:    s_add_u32 s24, s24, 3
; GFX9-NEXT:    s_addc_u32 s25, s25, 0
; GFX9-NEXT:    s_add_u32 s22, s22, 3
; GFX9-NEXT:    s_addc_u32 s23, s23, 0
; GFX9-NEXT:    s_add_u32 s20, s20, 3
; GFX9-NEXT:    s_addc_u32 s21, s21, 0
; GFX9-NEXT:    s_add_u32 s18, s18, 3
; GFX9-NEXT:    s_addc_u32 s19, s19, 0
; GFX9-NEXT:    s_add_u32 s16, s16, 3
; GFX9-NEXT:    s_addc_u32 s17, s17, 0
; GFX9-NEXT:    s_add_u32 s7, s7, 3
; GFX9-NEXT:    s_addc_u32 s6, s6, 0
; GFX9-NEXT:    s_add_u32 s9, s9, 3
; GFX9-NEXT:    s_addc_u32 s8, s8, 0
; GFX9-NEXT:    s_add_u32 s11, s11, 3
; GFX9-NEXT:    s_addc_u32 s10, s10, 0
; GFX9-NEXT:    s_add_u32 s13, s13, 3
; GFX9-NEXT:    s_addc_u32 s12, s12, 0
; GFX9-NEXT:    s_add_u32 s15, s15, 3
; GFX9-NEXT:    s_addc_u32 s14, s14, 0
; GFX9-NEXT:    s_add_u32 s41, s41, 3
; GFX9-NEXT:    s_addc_u32 s40, s40, 0
; GFX9-NEXT:    s_add_u32 s43, s43, 3
; GFX9-NEXT:    s_addc_u32 s42, s42, 0
; GFX9-NEXT:    s_add_u32 s45, s45, 3
; GFX9-NEXT:    s_addc_u32 s44, s44, 0
; GFX9-NEXT:  .LBB69_3: ; %end
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s47
; GFX9-NEXT:    v_mov_b32_e32 v15, s46
; GFX9-NEXT:    v_mov_b32_e32 v16, s45
; GFX9-NEXT:    v_mov_b32_e32 v17, s44
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v19, s42
; GFX9-NEXT:    v_mov_b32_e32 v20, s41
; GFX9-NEXT:    v_mov_b32_e32 v21, s40
; GFX9-NEXT:    v_mov_b32_e32 v22, s15
; GFX9-NEXT:    v_mov_b32_e32 v23, s14
; GFX9-NEXT:    v_mov_b32_e32 v24, s13
; GFX9-NEXT:    v_mov_b32_e32 v25, s12
; GFX9-NEXT:    v_mov_b32_e32 v26, s11
; GFX9-NEXT:    v_mov_b32_e32 v27, s10
; GFX9-NEXT:    v_mov_b32_e32 v28, s9
; GFX9-NEXT:    v_mov_b32_e32 v29, s8
; GFX9-NEXT:    v_mov_b32_e32 v30, s7
; GFX9-NEXT:    v_mov_b32_e32 v31, s6
; GFX9-NEXT:    s_setpc_b64 s[30:31]
; GFX9-NEXT:  .LBB69_4:
; GFX9-NEXT:    s_branch .LBB69_2
;
; GFX11-LABEL: bitcast_v16i64_to_v64i16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    v_readfirstlane_b32 s42, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_readfirstlane_b32 s5, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s7, v10
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_readfirstlane_b32 s11, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s13, v4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s15, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_readfirstlane_b32 s41, v0
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 s42, 0
; GFX11-NEXT:    s_cbranch_scc0 .LBB69_4
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s42
; GFX11-NEXT:    s_cbranch_vccnz .LBB69_3
; GFX11-NEXT:  .LBB69_2: ; %cmp.true
; GFX11-NEXT:    s_add_u32 s26, s26, 3
; GFX11-NEXT:    s_addc_u32 s27, s27, 0
; GFX11-NEXT:    s_add_u32 s24, s24, 3
; GFX11-NEXT:    s_addc_u32 s25, s25, 0
; GFX11-NEXT:    s_add_u32 s22, s22, 3
; GFX11-NEXT:    s_addc_u32 s23, s23, 0
; GFX11-NEXT:    s_add_u32 s20, s20, 3
; GFX11-NEXT:    s_addc_u32 s21, s21, 0
; GFX11-NEXT:    s_add_u32 s18, s18, 3
; GFX11-NEXT:    s_addc_u32 s19, s19, 0
; GFX11-NEXT:    s_add_u32 s16, s16, 3
; GFX11-NEXT:    s_addc_u32 s17, s17, 0
; GFX11-NEXT:    s_add_u32 s2, s2, 3
; GFX11-NEXT:    s_addc_u32 s3, s3, 0
; GFX11-NEXT:    s_add_u32 s0, s0, 3
; GFX11-NEXT:    s_addc_u32 s1, s1, 0
; GFX11-NEXT:    s_add_u32 s5, s5, 3
; GFX11-NEXT:    s_addc_u32 s4, s4, 0
; GFX11-NEXT:    s_add_u32 s7, s7, 3
; GFX11-NEXT:    s_addc_u32 s6, s6, 0
; GFX11-NEXT:    s_add_u32 s9, s9, 3
; GFX11-NEXT:    s_addc_u32 s8, s8, 0
; GFX11-NEXT:    s_add_u32 s11, s11, 3
; GFX11-NEXT:    s_addc_u32 s10, s10, 0
; GFX11-NEXT:    s_add_u32 s13, s13, 3
; GFX11-NEXT:    s_addc_u32 s12, s12, 0
; GFX11-NEXT:    s_add_u32 s15, s15, 3
; GFX11-NEXT:    s_addc_u32 s14, s14, 0
; GFX11-NEXT:    s_add_u32 s41, s41, 3
; GFX11-NEXT:    s_addc_u32 s40, s40, 0
; GFX11-NEXT:    s_add_u32 s28, s28, 3
; GFX11-NEXT:    s_addc_u32 s29, s29, 0
; GFX11-NEXT:  .LBB69_3: ; %end
; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:    v_dual_mov_b32 v16, s28 :: v_dual_mov_b32 v17, s29
; GFX11-NEXT:    v_dual_mov_b32 v18, s41 :: v_dual_mov_b32 v19, s40
; GFX11-NEXT:    v_dual_mov_b32 v20, s15 :: v_dual_mov_b32 v21, s14
; GFX11-NEXT:    v_dual_mov_b32 v22, s13 :: v_dual_mov_b32 v23, s12
; GFX11-NEXT:    v_dual_mov_b32 v24, s11 :: v_dual_mov_b32 v25, s10
; GFX11-NEXT:    v_dual_mov_b32 v26, s9 :: v_dual_mov_b32 v27, s8
; GFX11-NEXT:    v_dual_mov_b32 v28, s7 :: v_dual_mov_b32 v29, s6
; GFX11-NEXT:    v_dual_mov_b32 v30, s5 :: v_dual_mov_b32 v31, s4
; GFX11-NEXT:    s_setpc_b64 s[30:31]
; GFX11-NEXT:  .LBB69_4:
; GFX11-NEXT:    s_branch .LBB69_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <16 x i64> %a, splat (i64 3)
  %a2 = bitcast <16 x i64> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <16 x i64> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define <16 x i64> @bitcast_v64i16_to_v16i64(<64 x i16> %a, i32 %b) {
; SI-LABEL: bitcast_v64i16_to_v16i64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v19
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v20
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v21
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v22
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v23
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v24
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v25
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v27
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v17
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v61
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v23
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v22
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v21
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v19
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v18
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v17
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v10
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v62
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v28
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v29
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v30
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v63
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB70_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_or_b32_e32 v8, v8, v32
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    v_or_b32_e32 v0, v0, v49
; SI-NEXT:    v_or_b32_e32 v1, v1, v42
; SI-NEXT:    v_or_b32_e32 v2, v2, v34
; SI-NEXT:    v_or_b32_e32 v3, v3, v41
; SI-NEXT:    v_or_b32_e32 v4, v4, v48
; SI-NEXT:    v_or_b32_e32 v5, v5, v40
; SI-NEXT:    v_or_b32_e32 v6, v6, v33
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v54
; SI-NEXT:    v_or_b32_e32 v10, v10, v39
; SI-NEXT:    v_or_b32_e32 v11, v11, v53
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v52
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v37
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v15, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v17, v50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v18, v35
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v63
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB70_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB70_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v46
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v32, v8
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v62
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v61
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v60
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v59
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v58
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v57
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v56
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v47
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v44
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    s_mov_b32 s6, 0x30000
; SI-NEXT:    v_or_b32_e32 v2, v34, v2
; SI-NEXT:    v_or_b32_e32 v3, v41, v3
; SI-NEXT:    v_or_b32_e32 v4, v48, v4
; SI-NEXT:    v_or_b32_e32 v5, v40, v5
; SI-NEXT:    v_or_b32_e32 v6, v33, v6
; SI-NEXT:    v_or_b32_e32 v7, v55, v7
; SI-NEXT:    v_or_b32_e32 v9, v54, v9
; SI-NEXT:    v_or_b32_e32 v10, v39, v10
; SI-NEXT:    v_or_b32_e32 v11, v53, v11
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v38, v12
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v52, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v37, v14
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v51, v15
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v36, v16
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v50, v17
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v35, v18
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
; SI-NEXT:  .LBB70_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v16i64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB70_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v33, 3
; VI-NEXT:    v_add_u16_e32 v32, 3, v15
; VI-NEXT:    v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v32, v15
; VI-NEXT:    v_add_u16_e32 v32, 3, v14
; VI-NEXT:    v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v32, v14
; VI-NEXT:    v_add_u16_e32 v32, 3, v13
; VI-NEXT:    v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v32, v13
; VI-NEXT:    v_add_u16_e32 v32, 3, v12
; VI-NEXT:    v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v32, v12
; VI-NEXT:    v_add_u16_e32 v32, 3, v11
; VI-NEXT:    v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v32, v11
; VI-NEXT:    v_add_u16_e32 v32, 3, v10
; VI-NEXT:    v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v32, v10
; VI-NEXT:    v_add_u16_e32 v32, 3, v9
; VI-NEXT:    v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v32, v9
; VI-NEXT:    v_add_u16_e32 v32, 3, v8
; VI-NEXT:    v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v32, v8
; VI-NEXT:    v_add_u16_e32 v32, 3, v7
; VI-NEXT:    v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v32, v7
; VI-NEXT:    v_add_u16_e32 v32, 3, v6
; VI-NEXT:    v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v32, v6
; VI-NEXT:    v_add_u16_e32 v32, 3, v5
; VI-NEXT:    v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v32, v5
; VI-NEXT:    v_add_u16_e32 v32, 3, v4
; VI-NEXT:    v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v32, v4
; VI-NEXT:    v_add_u16_e32 v32, 3, v3
; VI-NEXT:    v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v32, v3
; VI-NEXT:    v_add_u16_e32 v32, 3, v2
; VI-NEXT:    v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v32, v2
; VI-NEXT:    v_add_u16_e32 v32, 3, v1
; VI-NEXT:    v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v32, v1
; VI-NEXT:    v_add_u16_e32 v32, 3, v0
; VI-NEXT:    v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v32, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v31
; VI-NEXT:    v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:    v_add_u16_e32 v32, 3, v30
; VI-NEXT:    v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v32, v30
; VI-NEXT:    v_add_u16_e32 v32, 3, v29
; VI-NEXT:    v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v32, v29
; VI-NEXT:    v_add_u16_e32 v32, 3, v28
; VI-NEXT:    v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v32, v28
; VI-NEXT:    v_add_u16_e32 v32, 3, v27
; VI-NEXT:    v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v32, v27
; VI-NEXT:    v_add_u16_e32 v32, 3, v26
; VI-NEXT:    v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v32, v26
; VI-NEXT:    v_add_u16_e32 v32, 3, v25
; VI-NEXT:    v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v32, v25
; VI-NEXT:    v_add_u16_e32 v32, 3, v24
; VI-NEXT:    v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v32, v24
; VI-NEXT:    v_add_u16_e32 v32, 3, v23
; VI-NEXT:    v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v32, v23
; VI-NEXT:    v_add_u16_e32 v32, 3, v22
; VI-NEXT:    v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v32, v22
; VI-NEXT:    v_add_u16_e32 v32, 3, v21
; VI-NEXT:    v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v32, v21
; VI-NEXT:    v_add_u16_e32 v32, 3, v20
; VI-NEXT:    v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v32, v20
; VI-NEXT:    v_add_u16_e32 v32, 3, v19
; VI-NEXT:    v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v32, v19
; VI-NEXT:    v_add_u16_e32 v32, 3, v18
; VI-NEXT:    v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v32, v18
; VI-NEXT:    v_add_u16_e32 v32, 3, v17
; VI-NEXT:    v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB70_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64i16_to_v16i64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB70_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB70_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v16i64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB70_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:  .LBB70_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define inreg <16 x i64> @bitcast_v64i16_to_v16i64_scalar(<64 x i16> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64i16_to_v16i64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    v_writelane_b32 v32, s70, 22
; SI-NEXT:    v_writelane_b32 v32, s71, 23
; SI-NEXT:    v_writelane_b32 v32, s80, 24
; SI-NEXT:    v_writelane_b32 v32, s81, 25
; SI-NEXT:    v_writelane_b32 v32, s82, 26
; SI-NEXT:    v_writelane_b32 v32, s83, 27
; SI-NEXT:    v_writelane_b32 v32, s84, 28
; SI-NEXT:    v_writelane_b32 v32, s85, 29
; SI-NEXT:    v_writelane_b32 v32, s86, 30
; SI-NEXT:    v_writelane_b32 v32, s87, 31
; SI-NEXT:    v_writelane_b32 v32, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s9, v16
; SI-NEXT:    v_writelane_b32 v32, s97, 33
; SI-NEXT:    s_lshr_b32 s14, s9, 16
; SI-NEXT:    v_readfirstlane_b32 s13, v14
; SI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v32, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s7, v17
; SI-NEXT:    v_readfirstlane_b32 s11, v15
; SI-NEXT:    s_lshr_b32 s72, s13, 16
; SI-NEXT:    v_readfirstlane_b32 s15, v13
; SI-NEXT:    v_readfirstlane_b32 s73, v12
; SI-NEXT:    v_readfirstlane_b32 s75, v11
; SI-NEXT:    v_readfirstlane_b32 s77, v10
; SI-NEXT:    v_readfirstlane_b32 s79, v9
; SI-NEXT:    v_readfirstlane_b32 s89, v8
; SI-NEXT:    v_readfirstlane_b32 s91, v7
; SI-NEXT:    v_readfirstlane_b32 s94, v6
; SI-NEXT:    v_readfirstlane_b32 s31, v5
; SI-NEXT:    v_readfirstlane_b32 s68, v4
; SI-NEXT:    v_readfirstlane_b32 s80, v3
; SI-NEXT:    v_readfirstlane_b32 s83, v2
; SI-NEXT:    v_readfirstlane_b32 s86, v1
; SI-NEXT:    v_readfirstlane_b32 s97, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v33, s14, 0
; SI-NEXT:    v_writelane_b32 v32, s99, 35
; SI-NEXT:    s_lshr_b32 s92, s29, 16
; SI-NEXT:    s_lshr_b32 s95, s28, 16
; SI-NEXT:    s_lshr_b32 s34, s27, 16
; SI-NEXT:    s_lshr_b32 s69, s26, 16
; SI-NEXT:    s_lshr_b32 s71, s25, 16
; SI-NEXT:    s_lshr_b32 s81, s24, 16
; SI-NEXT:    s_lshr_b32 s84, s23, 16
; SI-NEXT:    s_lshr_b32 s87, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s6, s19, 16
; SI-NEXT:    s_lshr_b32 s8, s18, 16
; SI-NEXT:    s_lshr_b32 s10, s17, 16
; SI-NEXT:    s_lshr_b32 s12, s16, 16
; SI-NEXT:    s_lshr_b32 vcc_lo, s7, 16
; SI-NEXT:    s_lshr_b32 vcc_hi, s11, 16
; SI-NEXT:    s_lshr_b32 s63, s15, 16
; SI-NEXT:    s_lshr_b32 s62, s73, 16
; SI-NEXT:    s_lshr_b32 s61, s75, 16
; SI-NEXT:    s_lshr_b32 s60, s77, 16
; SI-NEXT:    s_lshr_b32 s78, s79, 16
; SI-NEXT:    s_lshr_b32 s88, s89, 16
; SI-NEXT:    s_lshr_b32 s90, s91, 16
; SI-NEXT:    s_lshr_b32 s93, s94, 16
; SI-NEXT:    s_lshr_b32 s30, s31, 16
; SI-NEXT:    s_lshr_b32 s35, s68, 16
; SI-NEXT:    s_lshr_b32 s70, s80, 16
; SI-NEXT:    s_lshr_b32 s82, s83, 16
; SI-NEXT:    s_lshr_b32 s85, s86, 16
; SI-NEXT:    s_lshr_b32 s96, s97, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v33, s72, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v33, s63, 2
; SI-NEXT:    v_writelane_b32 v33, s62, 3
; SI-NEXT:    s_cbranch_scc0 .LBB71_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s10, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s60, 16
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s61, 16
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s62, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s63, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB71_3
; SI-NEXT:  .LBB71_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_or_b32 s5, s10, s5
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s36, s4, 0x30000
; SI-NEXT:    s_add_i32 s37, s5, 0x30000
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s38, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s39, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s40, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s41, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s42, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s43, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s44, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s45, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s46, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s47, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s48, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s97, s97, 3
; SI-NEXT:    s_add_i32 s49, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s86, s86, 3
; SI-NEXT:    s_add_i32 s50, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s83, s83, 3
; SI-NEXT:    s_add_i32 s51, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s80, s80, 3
; SI-NEXT:    s_add_i32 s52, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s68, s68, 3
; SI-NEXT:    s_add_i32 s53, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s31, s31, 3
; SI-NEXT:    s_add_i32 s54, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s94, s94, 3
; SI-NEXT:    s_add_i32 s55, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s91, s91, 3
; SI-NEXT:    s_add_i32 s56, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s89, s89, 3
; SI-NEXT:    s_add_i32 s57, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s79, s79, 3
; SI-NEXT:    s_add_i32 s58, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s77, s77, 3
; SI-NEXT:    s_add_i32 s59, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s76, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s75, s75, 3
; SI-NEXT:    s_add_i32 s60, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s74, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s73, s73, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 3
; SI-NEXT:    s_add_i32 s61, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 2
; SI-NEXT:    s_add_i32 s62, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 1
; SI-NEXT:    s_add_i32 s63, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s64, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 0
; SI-NEXT:    s_add_i32 s65, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s66, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s67, s4, 0x30000
; SI-NEXT:  .LBB71_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:    v_readlane_b32 s99, v32, 35
; SI-NEXT:    v_readlane_b32 s98, v32, 34
; SI-NEXT:    v_readlane_b32 s97, v32, 33
; SI-NEXT:    v_readlane_b32 s96, v32, 32
; SI-NEXT:    v_readlane_b32 s87, v32, 31
; SI-NEXT:    v_readlane_b32 s86, v32, 30
; SI-NEXT:    v_readlane_b32 s85, v32, 29
; SI-NEXT:    v_readlane_b32 s84, v32, 28
; SI-NEXT:    v_readlane_b32 s83, v32, 27
; SI-NEXT:    v_readlane_b32 s82, v32, 26
; SI-NEXT:    v_readlane_b32 s81, v32, 25
; SI-NEXT:    v_readlane_b32 s80, v32, 24
; SI-NEXT:    v_readlane_b32 s71, v32, 23
; SI-NEXT:    v_readlane_b32 s70, v32, 22
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB71_4:
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB71_2
;
; VI-LABEL: bitcast_v64i16_to_v16i64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB71_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB71_3
; VI-NEXT:  .LBB71_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s5, s46, 3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s47, 3
; VI-NEXT:    s_add_i32 s46, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s29, 3
; VI-NEXT:    s_add_i32 s47, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s28, 3
; VI-NEXT:    s_add_i32 s29, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s28, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s27, 3
; VI-NEXT:    s_add_i32 s28, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s27, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s26, 3
; VI-NEXT:    s_add_i32 s27, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s26, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s25, 3
; VI-NEXT:    s_add_i32 s26, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s25, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s24, 3
; VI-NEXT:    s_add_i32 s25, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s24, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s23, 3
; VI-NEXT:    s_add_i32 s24, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s23, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s22, 3
; VI-NEXT:    s_add_i32 s23, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s22, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s21, 3
; VI-NEXT:    s_add_i32 s22, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s21, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s20, 3
; VI-NEXT:    s_add_i32 s21, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s20, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s19, 3
; VI-NEXT:    s_add_i32 s20, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s19, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s18, 3
; VI-NEXT:    s_add_i32 s19, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s18, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s17, 3
; VI-NEXT:    s_add_i32 s18, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s17, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s16, 3
; VI-NEXT:    s_add_i32 s17, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s16, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s6, 3
; VI-NEXT:    s_add_i32 s16, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s6, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s7, 3
; VI-NEXT:    s_add_i32 s6, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s7, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s8, 3
; VI-NEXT:    s_add_i32 s7, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s8, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s9, 3
; VI-NEXT:    s_add_i32 s8, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s9, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s10, 3
; VI-NEXT:    s_add_i32 s9, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s10, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s11, 3
; VI-NEXT:    s_add_i32 s10, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s11, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s12, 3
; VI-NEXT:    s_add_i32 s11, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s12, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s13, 3
; VI-NEXT:    s_add_i32 s12, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s13, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s14, 3
; VI-NEXT:    s_add_i32 s13, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s14, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s15, 3
; VI-NEXT:    s_add_i32 s14, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s15, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s40, 3
; VI-NEXT:    s_add_i32 s15, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s41, 3
; VI-NEXT:    s_add_i32 s40, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s42, 3
; VI-NEXT:    s_add_i32 s41, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s43, 3
; VI-NEXT:    s_add_i32 s42, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s44, 3
; VI-NEXT:    s_add_i32 s43, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s45, 3
; VI-NEXT:    s_add_i32 s44, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s45, s4, 0x30000
; VI-NEXT:  .LBB71_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB71_4:
; VI-NEXT:    s_branch .LBB71_2
;
; GFX9-LABEL: bitcast_v64i16_to_v16i64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB71_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB71_4
; GFX9-NEXT:  .LBB71_2: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB71_5
; GFX9-NEXT:  .LBB71_3:
; GFX9-NEXT:    s_branch .LBB71_2
; GFX9-NEXT:  .LBB71_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB71_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v16i64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB71_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB71_4
; GFX11-NEXT:  .LBB71_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_branch .LBB71_5
; GFX11-NEXT:  .LBB71_3:
; GFX11-NEXT:    s_branch .LBB71_2
; GFX11-NEXT:  .LBB71_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB71_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <16 x i64>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <16 x i64>
  br label %end

end:
  %phi = phi <16 x i64> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x i64> %phi
}

define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) {
; SI-LABEL: bitcast_v16f64_to_v128i8:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB72_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v33, v32, v31, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v32, v31, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v32, v31, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v39, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v49, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v52, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v54, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v40, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v42, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v44, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v46, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v32
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v38, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v51, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v55, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v41, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v45, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB72_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB72_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_f64 v[31:32], v[31:32], 1.0
; SI-NEXT:    v_add_f64 v[29:30], v[29:30], 1.0
; SI-NEXT:    v_alignbit_b32 v33, v32, v31, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v32, v31, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v32, v31, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 24
; SI-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v30, v29, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 24
; SI-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v28, v27, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 24
; SI-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v26, v25, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 24
; SI-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v24, v23, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 24
; SI-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v22, v21, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 24
; SI-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v20, v19, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 24
; SI-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v18, v17, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 24
; SI-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v16, v15, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 24
; SI-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v14, v13, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 24
; SI-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v12, v11, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 24
; SI-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v10, v9, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v8, v7, 8
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; SI-NEXT:    v_add_f64 v[1:2], v[1:2], 1.0
; SI-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
; SI-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; SI-NEXT:    v_alignbit_b32 v37, v6, v5, 24
; SI-NEXT:    v_alignbit_b32 v39, v6, v5, 16
; SI-NEXT:    v_alignbit_b32 v49, v6, v5, 8
; SI-NEXT:    v_alignbit_b32 v52, v4, v3, 24
; SI-NEXT:    v_alignbit_b32 v54, v4, v3, 16
; SI-NEXT:    v_alignbit_b32 v40, v4, v3, 8
; SI-NEXT:    v_alignbit_b32 v42, v2, v1, 24
; SI-NEXT:    v_alignbit_b32 v44, v2, v1, 16
; SI-NEXT:    v_alignbit_b32 v46, v2, v1, 8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v32
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v38, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 8, v14
; SI-NEXT:    v_lshrrev_b32_e32 v51, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v55, 8, v12
; SI-NEXT:    v_lshrrev_b32_e32 v41, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v45, 8, v10
; SI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v57, 8, v8
; SI-NEXT:    v_lshrrev_b32_e32 v58, 24, v6
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v60, 8, v6
; SI-NEXT:    v_lshrrev_b32_e32 v61, 24, v4
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v63, 8, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; SI-NEXT:  .LBB72_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v46
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_and_b32_e32 v44, 0xff, v44
; SI-NEXT:    v_or_b32_e32 v1, v1, v46
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v42
; SI-NEXT:    v_or_b32_e32 v42, v42, v44
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v42
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v35
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v34
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v33
; SI-NEXT:    v_or_b32_e32 v2, v33, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v40
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v52
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v63
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v61
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v49
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v39
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v37
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v60
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v58
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v7
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v57
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v47
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v9
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v45
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v43
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v41
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v11
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v55
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v53
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v51
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v13
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v50
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v48
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v38
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v15
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v17
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v27
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v28
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v29
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v30
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v128i8:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; kill: killed $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB72_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[7:8]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    v_lshrrev_b32_e32 v56, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v57, 8, v11
; VI-NEXT:    v_lshrrev_b32_e32 v49, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v50, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v59, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v61, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v37, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v52, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v43, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v38, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v34, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v45, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v46, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v48, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v54, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v47, 8, v1
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:  .LBB72_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB72_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    v_add_f64 v[31:32], v[31:32], 1.0
; VI-NEXT:    v_add_f64 v[29:30], v[29:30], 1.0
; VI-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
; VI-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
; VI-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
; VI-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
; VI-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
; VI-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; VI-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; VI-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; VI-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; VI-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; VI-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; VI-NEXT:    v_add_f64 v[1:2], v[1:2], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    v_lshrrev_b32_e32 v42, 24, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[7:8]
; VI-NEXT:    v_lshrrev_b32_e32 v56, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v57, 8, v11
; VI-NEXT:    v_lshrrev_b32_e32 v49, 24, v10
; VI-NEXT:    v_lshrrev_b32_e32 v35, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v10
; VI-NEXT:    v_lshrrev_b32_e32 v36, 16, v9
; VI-NEXT:    v_lshrrev_b32_e32 v50, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v59, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v61, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v51, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v37, 24, v6
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v52, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v43, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v38, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v34, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v45, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v46, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v48, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v54, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v47, 8, v1
; VI-NEXT:  .LBB72_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v55, v55, v41, s4
; VI-NEXT:    v_perm_b32 v1, v1, v47, s4
; VI-NEXT:    v_perm_b32 v2, v2, v54, s4
; VI-NEXT:    v_lshlrev_b32_e32 v54, 16, v55
; VI-NEXT:    v_or_b32_e32 v1, v1, v54
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v48, v46, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v53, v40, s4
; VI-NEXT:    v_perm_b32 v2, v3, v45, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v44, v38, s4
; VI-NEXT:    v_perm_b32 v1, v4, v34, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v52, v39, s4
; VI-NEXT:    v_perm_b32 v2, v5, v43, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v62, v37, s4
; VI-NEXT:    v_perm_b32 v1, v6, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v7, v51, s4
; VI-NEXT:    v_perm_b32 v1, v33, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v60, v59, s4
; VI-NEXT:    v_perm_b32 v1, v8, v61, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v9, v50, s4
; VI-NEXT:    v_perm_b32 v1, v36, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v35, v49, s4
; VI-NEXT:    v_perm_b32 v1, v10, v58, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    v_perm_b32 v2, v11, v57, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v12, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v13, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v14, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v15, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v16, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v17, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v18, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v19, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v20, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v21, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v22, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v23, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v24, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v25, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v26, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v27, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v28, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v56, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v29, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v30, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v31, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v32, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v42, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v128i8:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    s_waitcnt vmcnt(41)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB72_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(41)
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 24, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    v_mov_b32_e32 v55, v39
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v8
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v8
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 8, v11
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 8, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 24, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 16, v32
; GFX9-NEXT:  .LBB72_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB72_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(40)
; GFX9-NEXT:    v_add_f64 v[31:32], v[31:32], 1.0
; GFX9-NEXT:    v_add_f64 v[29:30], v[29:30], 1.0
; GFX9-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
; GFX9-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
; GFX9-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
; GFX9-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
; GFX9-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
; GFX9-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX9-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; GFX9-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; GFX9-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; GFX9-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; GFX9-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; GFX9-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; GFX9-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; GFX9-NEXT:    v_add_f64 v[1:2], v[1:2], 1.0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX9-NEXT:    v_lshrrev_b64 v[39:40], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v8
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 24, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 16, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 8, v11
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v10
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 8, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v8
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 24, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v1
; GFX9-NEXT:  .LBB72_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v5, v5, v52, s4
; GFX9-NEXT:    v_perm_b32 v8, v8, v51, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v10, v10, v50, s4
; GFX9-NEXT:    v_perm_b32 v34, v34, v39, s4
; GFX9-NEXT:    v_perm_b32 v12, v12, v36, s4
; GFX9-NEXT:    v_perm_b32 v35, v35, v41, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v57, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v35, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, v2, v53, s4
; GFX9-NEXT:    v_perm_b32 v53, v47, v40, s4
; GFX9-NEXT:    v_perm_b32 v3, v3, v56, s4
; GFX9-NEXT:    v_perm_b32 v4, v4, v46, s4
; GFX9-NEXT:    v_perm_b32 v6, v6, v45, s4
; GFX9-NEXT:    v_perm_b32 v38, v38, v54, s4
; GFX9-NEXT:    v_perm_b32 v7, v7, v44, s4
; GFX9-NEXT:    v_perm_b32 v9, v9, v43, s4
; GFX9-NEXT:    v_perm_b32 v11, v11, v62, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v33, v33, v51, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v39, v49, v50, s4
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v36, v36, v49, s4
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v49, s4
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v14, v49, s4
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v49, v50, v49, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v15, v15, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v50, v51, v50, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v17, v51, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v51, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v51, v52, v51, s4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX9-NEXT:    v_perm_b32 v1, v63, v37, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshl_or_b32 v1, v53, 16, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    v_perm_b32 v1, v61, v60, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshl_or_b32 v1, v34, 16, v5
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    v_perm_b32 v1, v48, v59, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v6
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshl_or_b32 v1, v38, 16, v7
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v8
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshl_or_b32 v1, v33, 16, v9
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v10
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshl_or_b32 v1, v39, 16, v11
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v12
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshl_or_b32 v1, v36, 16, v13
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v14
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshl_or_b32 v1, v49, 16, v15
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v16
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshl_or_b32 v1, v50, 16, v17
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v18
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v19, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v51, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v20, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v21, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v22, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v23, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v24, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v25, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v26, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v27, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v58, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v28, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v30, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v1, v31, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v2, v42, v55, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v32, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v16f64_to_v128i8:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x2
; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB72_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB72_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB72_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_add_f64 v[31:32], v[31:32], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[29:30], v[29:30], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
; GFX11-TRUE16-NEXT:    v_add_f64 v[1:2], v[1:2], 1.0
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 8, v1
; GFX11-TRUE16-NEXT:  .LBB72_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v1.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v2.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.l, v3.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v66, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v3, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.l, v4.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v65, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v6.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v39, 16, v2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v5.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.l, v8.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.l, v7.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v66, v66, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v147, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v54, v64, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v39, 16, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v9.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v64, v65, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v66, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v55, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v10.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v11.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v64, 16, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v54, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v52, v55, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v12.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v13.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v14.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v15.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v12, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v13, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v14, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v15, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v16.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v17.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v18.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v19.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v16, v117, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v17, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v18, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v19, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v20.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v21.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v22.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v23.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v20, v103, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v21, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v99, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v14, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v24.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v15, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v25.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v26.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v27.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v27, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v28.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v29.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v30.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v31.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v28, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v29, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v30, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v31, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v32, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v29
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v16f64_to_v128i8:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x2
; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:12
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB72_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:  .LBB72_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB72_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_f64 v[19:20], v[19:20], 1.0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_add_f64 v[31:32], v[31:32], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[17:18], v[17:18], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[29:30], v[29:30], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[15:16], v[15:16], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[27:28], v[27:28], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[13:14], v[13:14], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[25:26], v[25:26], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[11:12], v[11:12], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[5:6], v[5:6], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[23:24], v[23:24], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[9:10], v[9:10], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[3:4], v[3:4], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[21:22], v[21:22], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[7:8], v[7:8], 1.0
; GFX11-FAKE16-NEXT:    v_add_f64 v[1:2], v[1:2], 1.0
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 8, v1
; GFX11-FAKE16-NEXT:  .LBB72_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v74, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v73, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v72, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v61, v65, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v59, v58, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v43, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v55, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v57, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v55, v56, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v64, v46, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, v41, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v54, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v182, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v53, v180, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v177, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v54, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v55, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v64, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v65, 16, v8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v53, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v54, 16, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v13, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v164, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v14, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v162, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v15, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v151, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v16, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v149, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v17, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v146, v49, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v18, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v144, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v19, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v133, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v20, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v131, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v21, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v128, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v118, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v115, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v24, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v113, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v25, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v102, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v26, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v97, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v28, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v87, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v29, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v84, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v30, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v31, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v71, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v69, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x13 ; 80-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:88
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define inreg <128 x i8> @bitcast_v16f64_to_v128i8_scalar(<16 x double> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16f64_to_v128i8_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_writelane_b32 v63, s70, 22
; SI-NEXT:    v_writelane_b32 v63, s71, 23
; SI-NEXT:    v_writelane_b32 v63, s80, 24
; SI-NEXT:    v_writelane_b32 v63, s81, 25
; SI-NEXT:    v_writelane_b32 v63, s82, 26
; SI-NEXT:    v_writelane_b32 v63, s83, 27
; SI-NEXT:    v_writelane_b32 v63, s84, 28
; SI-NEXT:    v_writelane_b32 v63, s85, 29
; SI-NEXT:    v_writelane_b32 v63, s86, 30
; SI-NEXT:    v_writelane_b32 v63, s87, 31
; SI-NEXT:    v_writelane_b32 v63, s96, 32
; SI-NEXT:    v_writelane_b32 v63, s97, 33
; SI-NEXT:    v_writelane_b32 v63, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s44, v19
; SI-NEXT:    v_writelane_b32 v63, s99, 35
; SI-NEXT:    v_readfirstlane_b32 s5, v18
; SI-NEXT:    v_readfirstlane_b32 s4, v17
; SI-NEXT:    v_readfirstlane_b32 s7, v16
; SI-NEXT:    v_readfirstlane_b32 s6, v15
; SI-NEXT:    v_readfirstlane_b32 s9, v14
; SI-NEXT:    v_readfirstlane_b32 s8, v13
; SI-NEXT:    v_readfirstlane_b32 s11, v12
; SI-NEXT:    v_readfirstlane_b32 s10, v11
; SI-NEXT:    v_readfirstlane_b32 s13, v10
; SI-NEXT:    v_readfirstlane_b32 s12, v9
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_readfirstlane_b32 s14, v7
; SI-NEXT:    v_readfirstlane_b32 s41, v6
; SI-NEXT:    v_readfirstlane_b32 s40, v5
; SI-NEXT:    v_readfirstlane_b32 s43, v4
; SI-NEXT:    v_readfirstlane_b32 s42, v3
; SI-NEXT:    v_readfirstlane_b32 s45, v2
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v1
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; SI-NEXT:    ; implicit-def: $vgpr61 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB73_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s46, s5, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 34
; SI-NEXT:    s_lshr_b32 s46, s5, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 35
; SI-NEXT:    s_lshr_b32 s46, s5, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 36
; SI-NEXT:    s_lshr_b32 s46, s7, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 37
; SI-NEXT:    s_lshr_b32 s46, s7, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 38
; SI-NEXT:    s_lshr_b32 s46, s7, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 39
; SI-NEXT:    s_lshr_b32 s46, s9, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 40
; SI-NEXT:    s_lshr_b32 s46, s9, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 41
; SI-NEXT:    s_lshr_b32 s46, s9, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 42
; SI-NEXT:    s_lshr_b32 s46, s11, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 43
; SI-NEXT:    s_lshr_b32 s46, s11, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 44
; SI-NEXT:    s_lshr_b32 s46, s11, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 45
; SI-NEXT:    s_lshr_b32 s46, s13, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 46
; SI-NEXT:    s_lshr_b32 s46, s13, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 47
; SI-NEXT:    s_lshr_b32 s46, s13, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 48
; SI-NEXT:    s_lshr_b32 s46, s15, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 49
; SI-NEXT:    s_lshr_b32 s46, s15, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 50
; SI-NEXT:    s_lshr_b32 s46, s15, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 51
; SI-NEXT:    s_lshr_b32 s46, s41, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 52
; SI-NEXT:    s_lshr_b32 s46, s41, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 53
; SI-NEXT:    s_lshr_b32 s46, s41, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 54
; SI-NEXT:    s_lshr_b32 s46, s43, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 55
; SI-NEXT:    s_lshr_b32 s46, s43, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 56
; SI-NEXT:    s_lshr_b32 s46, s43, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 57
; SI-NEXT:    s_lshr_b32 s46, s45, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 58
; SI-NEXT:    s_lshr_b32 s46, s45, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 59
; SI-NEXT:    s_lshr_b32 s46, s45, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 60
; SI-NEXT:    s_lshr_b32 s46, s29, 24
; SI-NEXT:    v_writelane_b32 v62, s46, 61
; SI-NEXT:    s_lshr_b32 s46, s29, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 62
; SI-NEXT:    s_lshr_b32 s46, s29, 8
; SI-NEXT:    v_writelane_b32 v62, s46, 63
; SI-NEXT:    s_lshr_b32 s46, s27, 24
; SI-NEXT:    v_writelane_b32 v61, s46, 0
; SI-NEXT:    s_lshr_b32 s46, s27, 16
; SI-NEXT:    v_writelane_b32 v61, s46, 1
; SI-NEXT:    s_lshr_b32 s46, s27, 8
; SI-NEXT:    v_writelane_b32 v61, s46, 2
; SI-NEXT:    s_lshr_b32 s46, s25, 24
; SI-NEXT:    v_writelane_b32 v61, s46, 3
; SI-NEXT:    s_lshr_b32 s46, s25, 16
; SI-NEXT:    v_writelane_b32 v61, s46, 4
; SI-NEXT:    s_lshr_b32 s46, s25, 8
; SI-NEXT:    v_writelane_b32 v61, s46, 5
; SI-NEXT:    s_lshr_b32 s46, s23, 24
; SI-NEXT:    v_writelane_b32 v61, s46, 6
; SI-NEXT:    s_lshr_b32 s46, s23, 16
; SI-NEXT:    v_writelane_b32 v61, s46, 7
; SI-NEXT:    s_lshr_b32 s46, s23, 8
; SI-NEXT:    v_writelane_b32 v61, s46, 8
; SI-NEXT:    s_lshr_b32 s46, s21, 24
; SI-NEXT:    v_writelane_b32 v61, s46, 9
; SI-NEXT:    s_lshr_b32 s46, s21, 16
; SI-NEXT:    v_writelane_b32 v61, s46, 10
; SI-NEXT:    s_lshr_b32 s46, s21, 8
; SI-NEXT:    v_writelane_b32 v61, s46, 11
; SI-NEXT:    s_lshr_b32 s46, s19, 24
; SI-NEXT:    v_writelane_b32 v61, s46, 12
; SI-NEXT:    s_lshr_b32 s46, s19, 16
; SI-NEXT:    v_writelane_b32 v61, s46, 13
; SI-NEXT:    s_lshr_b32 s46, s19, 8
; SI-NEXT:    v_writelane_b32 v61, s46, 14
; SI-NEXT:    s_lshr_b32 s46, s17, 24
; SI-NEXT:    v_writelane_b32 v61, s46, 15
; SI-NEXT:    s_lshr_b32 s46, s17, 16
; SI-NEXT:    v_writelane_b32 v61, s46, 16
; SI-NEXT:    s_lshr_b32 s46, s17, 8
; SI-NEXT:    v_writelane_b32 v61, s46, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; SI-NEXT:    v_writelane_b32 v62, s46, 32
; SI-NEXT:    v_writelane_b32 v62, s47, 33
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 8
; SI-NEXT:    v_writelane_b32 v62, s46, 30
; SI-NEXT:    v_writelane_b32 v62, s47, 31
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 24
; SI-NEXT:    v_writelane_b32 v62, s46, 28
; SI-NEXT:    v_writelane_b32 v62, s47, 29
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 16
; SI-NEXT:    v_writelane_b32 v62, s46, 26
; SI-NEXT:    v_writelane_b32 v62, s47, 27
; SI-NEXT:    s_lshr_b64 s[46:47], s[6:7], 8
; SI-NEXT:    v_writelane_b32 v62, s46, 24
; SI-NEXT:    v_writelane_b32 v62, s47, 25
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 24
; SI-NEXT:    v_writelane_b32 v62, s46, 22
; SI-NEXT:    v_writelane_b32 v62, s47, 23
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 16
; SI-NEXT:    v_writelane_b32 v62, s46, 20
; SI-NEXT:    v_writelane_b32 v62, s47, 21
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 8
; SI-NEXT:    v_writelane_b32 v62, s46, 18
; SI-NEXT:    v_writelane_b32 v62, s47, 19
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v62, s46, 16
; SI-NEXT:    v_writelane_b32 v62, s47, 17
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v62, s46, 14
; SI-NEXT:    v_writelane_b32 v62, s47, 15
; SI-NEXT:    s_lshr_b64 s[46:47], s[10:11], 8
; SI-NEXT:    v_writelane_b32 v62, s46, 12
; SI-NEXT:    v_writelane_b32 v62, s47, 13
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 24
; SI-NEXT:    v_writelane_b32 v62, s46, 10
; SI-NEXT:    v_writelane_b32 v62, s47, 11
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 16
; SI-NEXT:    v_writelane_b32 v62, s46, 8
; SI-NEXT:    v_writelane_b32 v62, s47, 9
; SI-NEXT:    s_lshr_b64 s[46:47], s[12:13], 8
; SI-NEXT:    v_writelane_b32 v62, s46, 6
; SI-NEXT:    v_writelane_b32 v62, s47, 7
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 24
; SI-NEXT:    v_writelane_b32 v62, s46, 4
; SI-NEXT:    v_writelane_b32 v62, s47, 5
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 16
; SI-NEXT:    v_writelane_b32 v62, s46, 2
; SI-NEXT:    v_writelane_b32 v62, s47, 3
; SI-NEXT:    s_lshr_b64 s[46:47], s[14:15], 8
; SI-NEXT:    v_writelane_b32 v62, s46, 0
; SI-NEXT:    s_lshr_b64 s[48:49], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v62, s47, 1
; SI-NEXT:    s_lshr_b64 s[50:51], s[40:41], 24
; SI-NEXT:    s_lshr_b64 s[52:53], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[54:55], s[40:41], 8
; SI-NEXT:    s_lshr_b64 s[64:65], s[42:43], 24
; SI-NEXT:    s_lshr_b64 s[66:67], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[68:69], s[42:43], 8
; SI-NEXT:    s_lshr_b64 s[70:71], s[44:45], 24
; SI-NEXT:    s_lshr_b64 s[80:81], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[84:85], s[28:29], 24
; SI-NEXT:    s_lshr_b64 s[86:87], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[96:97], s[28:29], 8
; SI-NEXT:    s_lshr_b64 s[98:99], s[26:27], 24
; SI-NEXT:    s_lshr_b64 s[46:47], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[26:27], 8
; SI-NEXT:    s_lshr_b64 s[58:59], s[24:25], 24
; SI-NEXT:    s_lshr_b64 s[60:61], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[24:25], 8
; SI-NEXT:    s_lshr_b64 s[72:73], s[22:23], 24
; SI-NEXT:    s_lshr_b64 s[74:75], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[22:23], 8
; SI-NEXT:    s_lshr_b64 s[78:79], s[20:21], 24
; SI-NEXT:    s_lshr_b64 s[88:89], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[20:21], 8
; SI-NEXT:    s_lshr_b64 s[92:93], s[18:19], 24
; SI-NEXT:    s_lshr_b64 s[94:95], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[18:19], 8
; SI-NEXT:    s_lshr_b64 s[34:35], s[16:17], 24
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_lshr_b64 s[38:39], s[16:17], 8
; SI-NEXT:    s_cbranch_execnz .LBB73_4
; SI-NEXT:  .LBB73_2: ; %cmp.true
; SI-NEXT:    v_add_f64 v[30:31], s[28:29], 1.0
; SI-NEXT:    v_add_f64 v[7:8], s[26:27], 1.0
; SI-NEXT:    v_lshr_b64 v[9:10], v[30:31], 24
; SI-NEXT:    v_lshr_b64 v[10:11], v[7:8], 24
; SI-NEXT:    v_add_f64 v[54:55], s[24:25], 1.0
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[54:55], 24
; SI-NEXT:    v_add_f64 v[42:43], s[22:23], 1.0
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[54:55], 16
; SI-NEXT:    v_add_f64 v[37:38], s[4:5], 1.0
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[42:43], 16
; SI-NEXT:    v_add_f64 v[56:57], s[20:21], 1.0
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[42:43], 8
; SI-NEXT:    v_lshr_b64 v[1:2], v[37:38], 24
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[56:57], 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[37:38], 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[56:57], 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[37:38], 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[56:57], 8
; SI-NEXT:    v_add_f64 v[35:36], s[6:7], 1.0
; SI-NEXT:    v_add_f64 v[19:20], s[18:19], 1.0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[35:36], 24
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[19:20], 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[35:36], 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[19:20], 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[35:36], 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[19:20], 8
; SI-NEXT:    v_add_f64 v[3:4], s[8:9], 1.0
; SI-NEXT:    v_add_f64 v[24:25], s[16:17], 1.0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[3:4], 24
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[24:25], 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[3:4], 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[24:25], 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[3:4], 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[24:25], 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f64 v[47:48], s[10:11], 1.0
; SI-NEXT:    v_add_f64 v[13:14], s[40:41], 1.0
; SI-NEXT:    v_lshr_b64 v[1:2], v[47:48], 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f64 v[32:33], s[14:15], 1.0
; SI-NEXT:    v_add_f64 v[28:29], s[42:43], 1.0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[47:48], 8
; SI-NEXT:    v_lshr_b64 v[17:18], v[13:14], 16
; SI-NEXT:    v_add_f64 v[21:22], s[44:45], 1.0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    v_lshr_b64 v[15:16], v[32:33], 8
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    v_lshr_b64 v[26:27], v[28:29], 8
; SI-NEXT:    v_readfirstlane_b32 s11, v48
; SI-NEXT:    v_lshr_b64 v[51:52], v[47:48], 16
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[1:2], v[28:29], 24
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshr_b64 v[16:17], v[28:29], 16
; SI-NEXT:    v_mov_b32_e32 v48, v28
; SI-NEXT:    v_lshr_b64 v[27:28], v[21:22], 8
; SI-NEXT:    v_readfirstlane_b32 s43, v29
; SI-NEXT:    v_lshr_b64 v[28:29], v[30:31], 16
; SI-NEXT:    v_readfirstlane_b32 s29, v31
; SI-NEXT:    v_lshr_b64 v[52:53], v[30:31], 8
; SI-NEXT:    v_readfirstlane_b32 s19, v20
; SI-NEXT:    v_readfirstlane_b32 s23, v43
; SI-NEXT:    v_readfirstlane_b32 s5, v38
; SI-NEXT:    v_lshr_b64 v[38:39], v[42:43], 24
; SI-NEXT:    v_add_f64 v[44:45], s[12:13], 1.0
; SI-NEXT:    v_readfirstlane_b32 s9, v4
; SI-NEXT:    v_lshr_b64 v[58:59], v[44:45], 24
; SI-NEXT:    v_lshr_b64 v[4:5], v[32:33], 16
; SI-NEXT:    v_readfirstlane_b32 s15, v33
; SI-NEXT:    v_readfirstlane_b32 s13, v45
; SI-NEXT:    v_readfirstlane_b32 s7, v36
; SI-NEXT:    v_lshr_b64 v[40:41], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[59:60], v[44:45], 8
; SI-NEXT:    v_lshr_b64 v[45:46], v[32:33], 24
; SI-NEXT:    v_lshr_b64 v[33:34], v[13:14], 24
; SI-NEXT:    v_lshr_b64 v[5:6], v[13:14], 8
; SI-NEXT:    v_mov_b32_e32 v36, v13
; SI-NEXT:    v_lshr_b64 v[12:13], v[21:22], 24
; SI-NEXT:    v_readfirstlane_b32 s41, v14
; SI-NEXT:    v_lshr_b64 v[13:14], v[7:8], 16
; SI-NEXT:    v_lshr_b64 v[49:50], v[54:55], 8
; SI-NEXT:    v_mov_b32_e32 v14, v9
; SI-NEXT:    v_readfirstlane_b32 s17, v25
; SI-NEXT:    v_mov_b32_e32 v25, v12
; SI-NEXT:    v_mov_b32_e32 v12, v5
; SI-NEXT:    v_mov_b32_e32 v5, v16
; SI-NEXT:    v_readfirstlane_b32 s27, v8
; SI-NEXT:    v_readfirstlane_b32 s45, v22
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[17:18], v[21:22], 16
; SI-NEXT:    v_lshr_b64 v[22:23], v[7:8], 8
; SI-NEXT:    v_mov_b32_e32 v23, v27
; SI-NEXT:    v_mov_b32_e32 v27, v17
; SI-NEXT:    v_mov_b32_e32 v17, v15
; SI-NEXT:    v_readfirstlane_b32 s21, v57
; SI-NEXT:    v_readfirstlane_b32 s25, v55
; SI-NEXT:    s_lshr_b32 s10, s5, 24
; SI-NEXT:    s_lshr_b32 s12, s5, 16
; SI-NEXT:    s_lshr_b32 s14, s5, 8
; SI-NEXT:    s_lshr_b32 s16, s7, 24
; SI-NEXT:    s_lshr_b32 s18, s7, 16
; SI-NEXT:    s_lshr_b32 s20, s7, 8
; SI-NEXT:    s_lshr_b32 s22, s9, 24
; SI-NEXT:    s_lshr_b32 s24, s9, 16
; SI-NEXT:    s_lshr_b32 s26, s9, 8
; SI-NEXT:    s_lshr_b32 s28, s11, 24
; SI-NEXT:    s_lshr_b32 s40, s11, 16
; SI-NEXT:    s_lshr_b32 s42, s11, 8
; SI-NEXT:    s_lshr_b32 s44, s13, 24
; SI-NEXT:    s_lshr_b32 s46, s13, 16
; SI-NEXT:    s_lshr_b32 s47, s13, 8
; SI-NEXT:    s_lshr_b32 s56, s15, 24
; SI-NEXT:    s_lshr_b32 s57, s15, 16
; SI-NEXT:    s_lshr_b32 s58, s15, 8
; SI-NEXT:    s_lshr_b32 s59, s41, 24
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mov_b32_e32 v29, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b32 s60, s41, 16
; SI-NEXT:    s_lshr_b32 s61, s41, 8
; SI-NEXT:    s_lshr_b32 s62, s43, 24
; SI-NEXT:    s_lshr_b32 s63, s43, 16
; SI-NEXT:    s_lshr_b32 s72, s43, 8
; SI-NEXT:    s_lshr_b32 s73, s45, 24
; SI-NEXT:    s_lshr_b32 s74, s45, 16
; SI-NEXT:    s_lshr_b32 s75, s45, 8
; SI-NEXT:    s_lshr_b32 s76, s29, 24
; SI-NEXT:    s_lshr_b32 s77, s29, 16
; SI-NEXT:    s_lshr_b32 s78, s29, 8
; SI-NEXT:    s_lshr_b32 s79, s27, 24
; SI-NEXT:    s_lshr_b32 s88, s27, 16
; SI-NEXT:    s_lshr_b32 s89, s27, 8
; SI-NEXT:    s_lshr_b32 s90, s25, 24
; SI-NEXT:    s_lshr_b32 s91, s25, 16
; SI-NEXT:    s_lshr_b32 s92, s25, 8
; SI-NEXT:    s_lshr_b32 s93, s23, 24
; SI-NEXT:    s_lshr_b32 s94, s23, 16
; SI-NEXT:    s_lshr_b32 s95, s23, 8
; SI-NEXT:    s_lshr_b32 s30, s21, 24
; SI-NEXT:    s_lshr_b32 s31, s21, 16
; SI-NEXT:    s_lshr_b32 s34, s21, 8
; SI-NEXT:    s_lshr_b32 s35, s19, 24
; SI-NEXT:    s_lshr_b32 s36, s19, 16
; SI-NEXT:    s_lshr_b32 s37, s19, 8
; SI-NEXT:    s_lshr_b32 s8, s17, 24
; SI-NEXT:    s_lshr_b32 vcc_lo, s17, 16
; SI-NEXT:    s_lshr_b32 s6, s17, 8
; SI-NEXT:    v_mov_b32_e32 v34, v4
; SI-NEXT:    v_mov_b32_e32 v53, v51
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v31, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v20, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v43, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v50, v10
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v46, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v16, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v8, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v15, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v51, v10
; SI-NEXT:    s_branch .LBB73_5
; SI-NEXT:  .LBB73_3:
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 0
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 1
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 2
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 3
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 4
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 5
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 6
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 7
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 8
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 9
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 10
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 11
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 12
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 13
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 14
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 15
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 16
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 17
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 18
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 19
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 20
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 21
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 22
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 23
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 24
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 25
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 26
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s49, 27
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 28
; SI-NEXT:    v_writelane_b32 v62, s49, 29
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 30
; SI-NEXT:    v_writelane_b32 v62, s49, 31
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    v_writelane_b32 v62, s48, 32
; SI-NEXT:    v_writelane_b32 v62, s49, 33
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    s_branch .LBB73_2
; SI-NEXT:  .LBB73_4:
; SI-NEXT:    v_mov_b32_e32 v1, s38
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s36
; SI-NEXT:    v_mov_b32_e32 v13, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 0
; SI-NEXT:    v_readlane_b32 s47, v62, 1
; SI-NEXT:    v_mov_b32_e32 v17, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 2
; SI-NEXT:    v_readlane_b32 s47, v62, 3
; SI-NEXT:    v_mov_b32_e32 v34, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 4
; SI-NEXT:    v_readlane_b32 s47, v62, 5
; SI-NEXT:    v_mov_b32_e32 v45, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 6
; SI-NEXT:    v_readlane_b32 s47, v62, 7
; SI-NEXT:    v_mov_b32_e32 v59, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 8
; SI-NEXT:    v_readlane_b32 s47, v62, 9
; SI-NEXT:    v_mov_b32_e32 v40, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 10
; SI-NEXT:    v_readlane_b32 s47, v62, 11
; SI-NEXT:    v_mov_b32_e32 v58, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 12
; SI-NEXT:    v_readlane_b32 s47, v62, 13
; SI-NEXT:    v_mov_b32_e32 v9, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 14
; SI-NEXT:    v_readlane_b32 s47, v62, 15
; SI-NEXT:    v_mov_b32_e32 v53, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 16
; SI-NEXT:    v_readlane_b32 s47, v62, 17
; SI-NEXT:    v_mov_b32_e32 v50, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 18
; SI-NEXT:    v_readlane_b32 s47, v62, 19
; SI-NEXT:    v_mov_b32_e32 v29, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 20
; SI-NEXT:    v_readlane_b32 s47, v62, 21
; SI-NEXT:    v_mov_b32_e32 v31, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 22
; SI-NEXT:    v_readlane_b32 s47, v62, 23
; SI-NEXT:    v_mov_b32_e32 v20, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 24
; SI-NEXT:    v_readlane_b32 s47, v62, 25
; SI-NEXT:    v_mov_b32_e32 v43, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 26
; SI-NEXT:    v_readlane_b32 s47, v62, 27
; SI-NEXT:    v_mov_b32_e32 v46, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 28
; SI-NEXT:    v_readlane_b32 s47, v62, 29
; SI-NEXT:    v_mov_b32_e32 v16, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 30
; SI-NEXT:    v_readlane_b32 s47, v62, 31
; SI-NEXT:    v_mov_b32_e32 v51, s46
; SI-NEXT:    v_readlane_b32 s46, v62, 32
; SI-NEXT:    v_mov_b32_e32 v3, s8
; SI-NEXT:    v_readlane_b32 s47, v62, 33
; SI-NEXT:    v_mov_b32_e32 v38, s72
; SI-NEXT:    v_mov_b32_e32 v49, s62
; SI-NEXT:    v_mov_b32_e32 v22, s56
; SI-NEXT:    v_mov_b32_e32 v24, s16
; SI-NEXT:    v_mov_b32_e32 v19, s18
; SI-NEXT:    v_mov_b32_e32 v56, s20
; SI-NEXT:    v_mov_b32_e32 v42, s22
; SI-NEXT:    v_mov_b32_e32 v54, s24
; SI-NEXT:    v_mov_b32_e32 v7, s26
; SI-NEXT:    v_mov_b32_e32 v30, s28
; SI-NEXT:    v_mov_b32_e32 v21, s44
; SI-NEXT:    v_mov_b32_e32 v48, s42
; SI-NEXT:    v_mov_b32_e32 v36, s40
; SI-NEXT:    v_mov_b32_e32 v32, s14
; SI-NEXT:    v_mov_b32_e32 v44, s12
; SI-NEXT:    v_mov_b32_e32 v47, s10
; SI-NEXT:    v_mov_b32_e32 v35, s6
; SI-NEXT:    v_mov_b32_e32 v37, s4
; SI-NEXT:    v_mov_b32_e32 v52, s96
; SI-NEXT:    v_mov_b32_e32 v28, s86
; SI-NEXT:    v_mov_b32_e32 v14, s84
; SI-NEXT:    v_mov_b32_e32 v23, s82
; SI-NEXT:    v_mov_b32_e32 v27, s80
; SI-NEXT:    v_mov_b32_e32 v25, s70
; SI-NEXT:    v_mov_b32_e32 v26, s68
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s34
; SI-NEXT:    v_mov_b32_e32 v5, s66
; SI-NEXT:    v_mov_b32_e32 v12, s54
; SI-NEXT:    v_mov_b32_e32 v33, s50
; SI-NEXT:    v_mov_b32_e32 v8, s48
; SI-NEXT:    v_mov_b32_e32 v15, s46
; SI-NEXT:    v_readlane_b32 s10, v62, 34
; SI-NEXT:    v_readlane_b32 s12, v62, 35
; SI-NEXT:    v_readlane_b32 s14, v62, 36
; SI-NEXT:    v_readlane_b32 s16, v62, 37
; SI-NEXT:    v_readlane_b32 s18, v62, 38
; SI-NEXT:    v_readlane_b32 s20, v62, 39
; SI-NEXT:    v_readlane_b32 s22, v62, 40
; SI-NEXT:    v_readlane_b32 s24, v62, 41
; SI-NEXT:    v_readlane_b32 s26, v62, 42
; SI-NEXT:    v_readlane_b32 s28, v62, 43
; SI-NEXT:    v_readlane_b32 s40, v62, 44
; SI-NEXT:    v_readlane_b32 s42, v62, 45
; SI-NEXT:    v_readlane_b32 s44, v62, 46
; SI-NEXT:    v_readlane_b32 s46, v62, 47
; SI-NEXT:    v_readlane_b32 s47, v62, 48
; SI-NEXT:    v_readlane_b32 s56, v62, 49
; SI-NEXT:    v_readlane_b32 s57, v62, 50
; SI-NEXT:    v_readlane_b32 s59, v62, 52
; SI-NEXT:    v_readlane_b32 s61, v62, 54
; SI-NEXT:    v_readlane_b32 s62, v62, 55
; SI-NEXT:    v_readlane_b32 s63, v62, 56
; SI-NEXT:    v_readlane_b32 s72, v62, 57
; SI-NEXT:    v_readlane_b32 s73, v62, 58
; SI-NEXT:    v_readlane_b32 s75, v62, 60
; SI-NEXT:    v_readlane_b32 s77, v62, 62
; SI-NEXT:    v_readlane_b32 s79, v61, 0
; SI-NEXT:    v_readlane_b32 s89, v61, 2
; SI-NEXT:    v_readlane_b32 s91, v61, 4
; SI-NEXT:    v_readlane_b32 s93, v61, 6
; SI-NEXT:    v_readlane_b32 s95, v61, 8
; SI-NEXT:    v_readlane_b32 s31, v61, 10
; SI-NEXT:    v_readlane_b32 s34, v61, 11
; SI-NEXT:    v_readlane_b32 s35, v61, 12
; SI-NEXT:    v_readlane_b32 s36, v61, 13
; SI-NEXT:    v_readlane_b32 s37, v61, 14
; SI-NEXT:    v_readlane_b32 s8, v61, 15
; SI-NEXT:    v_readlane_b32 vcc_lo, v61, 16
; SI-NEXT:    v_readlane_b32 s6, v61, 17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s30
; SI-NEXT:    v_readlane_b32 s30, v61, 9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s94
; SI-NEXT:    v_readlane_b32 s94, v61, 7
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s92
; SI-NEXT:    v_readlane_b32 s92, v61, 5
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s90
; SI-NEXT:    v_readlane_b32 s90, v61, 3
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s88
; SI-NEXT:    v_readlane_b32 s88, v61, 1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s78
; SI-NEXT:    v_readlane_b32 s78, v62, 63
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s76
; SI-NEXT:    v_readlane_b32 s76, v62, 61
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s74
; SI-NEXT:    v_readlane_b32 s74, v62, 59
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s60
; SI-NEXT:    v_readlane_b32 s60, v62, 53
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s58
; SI-NEXT:    v_readlane_b32 s58, v62, 51
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, s98
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s52
; SI-NEXT:    v_mov_b32_e32 v1, s64
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:  .LBB73_5: ; %end
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v24
; SI-NEXT:    s_and_b32 s4, s17, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, vcc_lo, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_lshl_b32 s6, s37, 8
; SI-NEXT:    s_lshl_b32 s8, s35, 24
; SI-NEXT:    v_readlane_b32 s99, v63, 35
; SI-NEXT:    v_readlane_b32 s98, v63, 34
; SI-NEXT:    v_readlane_b32 s97, v63, 33
; SI-NEXT:    v_readlane_b32 s96, v63, 32
; SI-NEXT:    v_readlane_b32 s87, v63, 31
; SI-NEXT:    v_readlane_b32 s86, v63, 30
; SI-NEXT:    v_readlane_b32 s85, v63, 29
; SI-NEXT:    v_readlane_b32 s84, v63, 28
; SI-NEXT:    v_readlane_b32 s83, v63, 27
; SI-NEXT:    v_readlane_b32 s82, v63, 26
; SI-NEXT:    v_readlane_b32 s81, v63, 25
; SI-NEXT:    v_readlane_b32 s80, v63, 24
; SI-NEXT:    v_readlane_b32 s71, v63, 23
; SI-NEXT:    v_readlane_b32 s70, v63, 22
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v10
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v19
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s4, s19, 0xff
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s36, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_lshl_b32 s6, s34, 8
; SI-NEXT:    s_lshl_b32 s8, s30, 24
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v18
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v56
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s4, s21, 0xff
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s31, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_lshl_b32 s6, s95, 8
; SI-NEXT:    s_lshl_b32 s8, s93, 24
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v55
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v55
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v55
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v42
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s4, s23, 0xff
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s94, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v38
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_lshl_b32 s6, s92, 8
; SI-NEXT:    s_lshl_b32 s8, s90, 24
; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v41
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v41
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v54
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v49
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    s_and_b32 s4, s25, 0xff
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s91, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_lshl_b32 s6, s89, 8
; SI-NEXT:    s_lshl_b32 s8, s79, 24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v38
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v7
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v22
; SI-NEXT:    s_and_b32 s4, s27, 0xff
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v13
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s88, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_lshl_b32 s6, s78, 8
; SI-NEXT:    s_lshl_b32 s8, s76, 24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v30
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v52
; SI-NEXT:    s_and_b32 s4, s29, 0xff
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v28
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s77, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v14
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 48, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v21
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v23
; SI-NEXT:    s_and_b32 s4, s45, 0xff
; SI-NEXT:    s_lshl_b32 s6, s75, 8
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v27
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s74, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v25
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s73, 24
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 56, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v48
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v26
; SI-NEXT:    s_and_b32 s4, s43, 0xff
; SI-NEXT:    s_lshl_b32 s6, s72, 8
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v5
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s63, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v1
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s62, 24
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 64, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v36
; SI-NEXT:    s_and_b32 s4, s41, 0xff
; SI-NEXT:    s_lshl_b32 s6, s61, 8
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s60, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v33
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s59, 24
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_lshl_b32 s6, s58, 8
; SI-NEXT:    s_lshl_b32 s8, s56, 24
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v20
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v37
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v32
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v17
; SI-NEXT:    s_and_b32 s4, s15, 0xff
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v34
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s57, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v45
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x50, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v44
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v59
; SI-NEXT:    s_and_b32 s4, s13, 0xff
; SI-NEXT:    s_lshl_b32 s6, s47, 8
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v40
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s46, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v58
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s44, 24
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x58, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v47
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v9
; SI-NEXT:    s_and_b32 s4, s11, 0xff
; SI-NEXT:    s_lshl_b32 s6, s42, 8
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v53
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s40, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v50
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s28, 24
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x60, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v29
; SI-NEXT:    s_and_b32 s4, s9, 0xff
; SI-NEXT:    s_lshl_b32 s6, s26, 8
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v31
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s24, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s22, 24
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x68, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    buffer_store_dword v4, v2, s[0:3], 0 offen
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v35
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v43
; SI-NEXT:    s_and_b32 s4, s7, 0xff
; SI-NEXT:    s_lshl_b32 s6, s20, 8
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v46
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s18, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v16
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s16, 24
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x70, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    v_mov_b32_e32 v3, s4
; SI-NEXT:    buffer_store_dword v3, v2, s[0:3], 0 offen
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v51
; SI-NEXT:    s_and_b32 s4, s5, 0xff
; SI-NEXT:    s_lshl_b32 s5, s14, 8
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s12, 0xff
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v15
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s6, s10, 24
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, s4
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v128i8_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v63, s30, 0
; VI-NEXT:    v_writelane_b32 v63, s31, 1
; VI-NEXT:    v_writelane_b32 v63, s34, 2
; VI-NEXT:    v_writelane_b32 v63, s35, 3
; VI-NEXT:    v_writelane_b32 v63, s36, 4
; VI-NEXT:    v_writelane_b32 v63, s37, 5
; VI-NEXT:    v_writelane_b32 v63, s38, 6
; VI-NEXT:    v_writelane_b32 v63, s39, 7
; VI-NEXT:    v_writelane_b32 v63, s48, 8
; VI-NEXT:    v_writelane_b32 v63, s49, 9
; VI-NEXT:    v_writelane_b32 v63, s50, 10
; VI-NEXT:    v_writelane_b32 v63, s51, 11
; VI-NEXT:    v_writelane_b32 v63, s52, 12
; VI-NEXT:    v_writelane_b32 v63, s53, 13
; VI-NEXT:    v_writelane_b32 v63, s54, 14
; VI-NEXT:    v_writelane_b32 v63, s55, 15
; VI-NEXT:    v_writelane_b32 v63, s64, 16
; VI-NEXT:    v_writelane_b32 v63, s65, 17
; VI-NEXT:    v_writelane_b32 v63, s66, 18
; VI-NEXT:    v_writelane_b32 v63, s67, 19
; VI-NEXT:    v_writelane_b32 v63, s68, 20
; VI-NEXT:    v_writelane_b32 v63, s69, 21
; VI-NEXT:    v_writelane_b32 v63, s70, 22
; VI-NEXT:    v_writelane_b32 v63, s71, 23
; VI-NEXT:    v_writelane_b32 v63, s80, 24
; VI-NEXT:    v_writelane_b32 v63, s81, 25
; VI-NEXT:    v_writelane_b32 v63, s82, 26
; VI-NEXT:    v_writelane_b32 v63, s83, 27
; VI-NEXT:    v_writelane_b32 v63, s84, 28
; VI-NEXT:    v_writelane_b32 v63, s85, 29
; VI-NEXT:    v_writelane_b32 v63, s86, 30
; VI-NEXT:    v_readfirstlane_b32 s44, v19
; VI-NEXT:    v_writelane_b32 v63, s87, 31
; VI-NEXT:    v_readfirstlane_b32 s5, v18
; VI-NEXT:    v_readfirstlane_b32 s4, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s6, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s8, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s10, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s12, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s14, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s40, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s42, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    s_cmp_lg_u32 s44, 0
; VI-NEXT:    v_readfirstlane_b32 s44, v1
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; VI-NEXT:    s_cbranch_scc0 .LBB73_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 0
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 1
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s41, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s41, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s43, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s43, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s29, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s27, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s25, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s23, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s23, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s23, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s21, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s21, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s19, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s19, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s19, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s17, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s17, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s17, 8
; VI-NEXT:    s_lshr_b32 s81, s14, 16
; VI-NEXT:    s_lshr_b32 s80, s14, 8
; VI-NEXT:    s_lshr_b32 s83, s40, 16
; VI-NEXT:    s_lshr_b32 s82, s40, 8
; VI-NEXT:    s_lshr_b32 s85, s42, 16
; VI-NEXT:    s_lshr_b32 s84, s42, 8
; VI-NEXT:    s_lshr_b32 s51, s44, 16
; VI-NEXT:    s_lshr_b32 s50, s44, 8
; VI-NEXT:    s_lshr_b32 s52, s28, 16
; VI-NEXT:    s_lshr_b32 s86, s28, 8
; VI-NEXT:    s_lshr_b32 s87, s26, 16
; VI-NEXT:    s_lshr_b32 s53, s26, 8
; VI-NEXT:    s_lshr_b32 s55, s24, 16
; VI-NEXT:    s_lshr_b32 s54, s24, 8
; VI-NEXT:    s_lshr_b32 s65, s22, 16
; VI-NEXT:    s_lshr_b32 s64, s22, 8
; VI-NEXT:    s_lshr_b32 s67, s20, 16
; VI-NEXT:    s_lshr_b32 s66, s20, 8
; VI-NEXT:    s_lshr_b32 s69, s18, 16
; VI-NEXT:    s_lshr_b32 s68, s18, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 57
; VI-NEXT:    s_lshr_b32 s71, s16, 16
; VI-NEXT:    s_lshr_b32 s70, s16, 8
; VI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:    s_cbranch_execnz .LBB73_4
; VI-NEXT:  .LBB73_2: ; %cmp.true
; VI-NEXT:    v_add_f64 v[13:14], s[4:5], 1.0
; VI-NEXT:    v_add_f64 v[1:2], s[6:7], 1.0
; VI-NEXT:    v_add_f64 v[3:4], s[8:9], 1.0
; VI-NEXT:    v_add_f64 v[5:6], s[10:11], 1.0
; VI-NEXT:    v_add_f64 v[7:8], s[12:13], 1.0
; VI-NEXT:    v_add_f64 v[9:10], s[14:15], 1.0
; VI-NEXT:    v_add_f64 v[11:12], s[40:41], 1.0
; VI-NEXT:    v_add_f64 v[15:16], s[42:43], 1.0
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    v_add_f64 v[17:18], s[44:45], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[1:2]
; VI-NEXT:    v_add_f64 v[19:20], s[28:29], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[3:4]
; VI-NEXT:    v_add_f64 v[21:22], s[26:27], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[5:6]
; VI-NEXT:    v_add_f64 v[23:24], s[24:25], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[7:8]
; VI-NEXT:    v_add_f64 v[25:26], s[22:23], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    v_add_f64 v[27:28], s[20:21], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    v_add_f64 v[29:30], s[18:19], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    v_add_f64 v[31:32], s[16:17], 1.0
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; VI-NEXT:    v_lshrrev_b32_e32 v51, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    v_lshrrev_b32_e32 v37, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    v_lshrrev_b64 v[42:43], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; VI-NEXT:    v_lshrrev_b32_e32 v53, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; VI-NEXT:    v_lshrrev_b32_e32 v38, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; VI-NEXT:    v_lshrrev_b32_e32 v47, 8, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v2
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v4
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v4
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v4
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v3
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v3
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v5
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v8
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v7
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    v_lshrrev_b32_e32 v34, 24, v24
; VI-NEXT:    v_lshrrev_b32_e32 v56, 16, v24
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v24
; VI-NEXT:    v_lshrrev_b32_e32 v54, 24, v26
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v26
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v39, 24, v28
; VI-NEXT:    v_lshrrev_b32_e32 v58, 16, v28
; VI-NEXT:    v_lshrrev_b32_e32 v49, 8, v28
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v27
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v27
; VI-NEXT:    v_lshrrev_b32_e32 v59, 24, v30
; VI-NEXT:    v_lshrrev_b32_e32 v40, 16, v30
; VI-NEXT:    v_lshrrev_b32_e32 v60, 8, v30
; VI-NEXT:    v_lshrrev_b32_e32 v43, 16, v29
; VI-NEXT:    v_lshrrev_b32_e32 v44, 8, v29
; VI-NEXT:    v_lshrrev_b32_e32 v61, 24, v32
; VI-NEXT:    v_lshrrev_b32_e32 v55, 16, v32
; VI-NEXT:    v_lshrrev_b32_e32 v36, 8, v32
; VI-NEXT:    v_lshrrev_b32_e32 v45, 16, v31
; VI-NEXT:    v_lshrrev_b32_e32 v46, 8, v31
; VI-NEXT:    s_branch .LBB73_5
; VI-NEXT:  .LBB73_3:
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr71
; VI-NEXT:    ; implicit-def: $sgpr68
; VI-NEXT:    ; implicit-def: $sgpr69
; VI-NEXT:    ; implicit-def: $sgpr66
; VI-NEXT:    ; implicit-def: $sgpr67
; VI-NEXT:    ; implicit-def: $sgpr64
; VI-NEXT:    ; implicit-def: $sgpr65
; VI-NEXT:    ; implicit-def: $sgpr54
; VI-NEXT:    ; implicit-def: $sgpr55
; VI-NEXT:    ; implicit-def: $sgpr53
; VI-NEXT:    ; implicit-def: $sgpr87
; VI-NEXT:    ; implicit-def: $sgpr86
; VI-NEXT:    ; implicit-def: $sgpr52
; VI-NEXT:    ; implicit-def: $sgpr50
; VI-NEXT:    ; implicit-def: $sgpr51
; VI-NEXT:    ; implicit-def: $sgpr84
; VI-NEXT:    ; implicit-def: $sgpr85
; VI-NEXT:    ; implicit-def: $sgpr82
; VI-NEXT:    ; implicit-def: $sgpr83
; VI-NEXT:    ; implicit-def: $sgpr80
; VI-NEXT:    ; implicit-def: $sgpr81
; VI-NEXT:    ; implicit-def: $sgpr48
; VI-NEXT:    ; implicit-def: $sgpr38
; VI-NEXT:    ; implicit-def: $sgpr36
; VI-NEXT:    ; implicit-def: $sgpr34
; VI-NEXT:    ; implicit-def: $sgpr30
; VI-NEXT:    ; implicit-def: $sgpr90
; VI-NEXT:    ; implicit-def: $sgpr88
; VI-NEXT:    ; implicit-def: $sgpr78
; VI-NEXT:    ; implicit-def: $sgpr76
; VI-NEXT:    ; implicit-def: $sgpr74
; VI-NEXT:    ; implicit-def: $sgpr72
; VI-NEXT:    ; implicit-def: $sgpr62
; VI-NEXT:    ; implicit-def: $sgpr60
; VI-NEXT:    ; implicit-def: $sgpr58
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    s_branch .LBB73_2
; VI-NEXT:  .LBB73_4:
; VI-NEXT:    v_mov_b32_e32 v42, s48
; VI-NEXT:    v_mov_b32_e32 v41, s38
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s36
; VI-NEXT:    v_mov_b32_e32 v34, s65
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s64
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s55
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s54
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s87
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s53
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s52
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s86
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s51
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s50
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s85
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s84
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s83
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s82
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s81
; VI-NEXT:    v_mov_b32_e32 v13, s4
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s80
; VI-NEXT:    v_readlane_b32 s4, v62, 0
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 1
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 2
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 3
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 4
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 5
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 6
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 7
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 8
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 9
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 10
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 11
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 12
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 13
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 14
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s34
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 15
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 16
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 17
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 18
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 19
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 20
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 21
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 22
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 23
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 24
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 25
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 26
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 27
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 28
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 29
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 30
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 31
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 32
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 33
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 34
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 35
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 36
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 37
; VI-NEXT:    v_mov_b32_e32 v51, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 38
; VI-NEXT:    v_mov_b32_e32 v37, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 39
; VI-NEXT:    v_mov_b32_e32 v52, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 40
; VI-NEXT:    v_mov_b32_e32 v53, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 41
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s30
; VI-NEXT:    v_mov_b32_e32 v38, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 42
; VI-NEXT:    v_mov_b32_e32 v47, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 43
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 44
; VI-NEXT:    v_mov_b32_e32 v56, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 45
; VI-NEXT:    v_mov_b32_e32 v48, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 46
; VI-NEXT:    v_mov_b32_e32 v54, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 47
; VI-NEXT:    v_mov_b32_e32 v57, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 48
; VI-NEXT:    v_mov_b32_e32 v35, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 49
; VI-NEXT:    v_mov_b32_e32 v39, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 50
; VI-NEXT:    v_mov_b32_e32 v58, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 51
; VI-NEXT:    v_mov_b32_e32 v49, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 52
; VI-NEXT:    v_mov_b32_e32 v59, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 53
; VI-NEXT:    v_mov_b32_e32 v40, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 54
; VI-NEXT:    v_mov_b32_e32 v60, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 55
; VI-NEXT:    v_mov_b32_e32 v61, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 56
; VI-NEXT:    v_mov_b32_e32 v55, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 57
; VI-NEXT:    v_mov_b32_e32 v14, s5
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_mov_b32_e32 v2, s7
; VI-NEXT:    v_mov_b32_e32 v3, s8
; VI-NEXT:    v_mov_b32_e32 v4, s9
; VI-NEXT:    v_mov_b32_e32 v5, s10
; VI-NEXT:    v_mov_b32_e32 v6, s11
; VI-NEXT:    v_mov_b32_e32 v7, s12
; VI-NEXT:    v_mov_b32_e32 v8, s13
; VI-NEXT:    v_mov_b32_e32 v9, s14
; VI-NEXT:    v_mov_b32_e32 v10, s15
; VI-NEXT:    v_mov_b32_e32 v11, s40
; VI-NEXT:    v_mov_b32_e32 v12, s41
; VI-NEXT:    v_mov_b32_e32 v15, s42
; VI-NEXT:    v_mov_b32_e32 v16, s43
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s45
; VI-NEXT:    v_mov_b32_e32 v19, s28
; VI-NEXT:    v_mov_b32_e32 v20, s29
; VI-NEXT:    v_mov_b32_e32 v21, s26
; VI-NEXT:    v_mov_b32_e32 v22, s27
; VI-NEXT:    v_mov_b32_e32 v23, s24
; VI-NEXT:    v_mov_b32_e32 v24, s25
; VI-NEXT:    v_mov_b32_e32 v25, s22
; VI-NEXT:    v_mov_b32_e32 v26, s23
; VI-NEXT:    v_mov_b32_e32 v27, s20
; VI-NEXT:    v_mov_b32_e32 v28, s21
; VI-NEXT:    v_mov_b32_e32 v29, s18
; VI-NEXT:    v_mov_b32_e32 v30, s19
; VI-NEXT:    v_mov_b32_e32 v31, s16
; VI-NEXT:    v_mov_b32_e32 v32, s17
; VI-NEXT:    v_mov_b32_e32 v45, s71
; VI-NEXT:    v_mov_b32_e32 v46, s70
; VI-NEXT:    v_mov_b32_e32 v43, s69
; VI-NEXT:    v_mov_b32_e32 v44, s68
; VI-NEXT:    v_mov_b32_e32 v50, s67
; VI-NEXT:    v_mov_b32_e32 v33, s66
; VI-NEXT:    v_mov_b32_e32 v36, s4
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s90
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s88
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s78
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s76
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s74
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s72
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s62
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s60
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s58
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s56
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v41, s46
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:  .LBB73_5: ; %end
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v41, v45, v42, s4
; VI-NEXT:    v_perm_b32 v31, v31, v46, s4
; VI-NEXT:    v_perm_b32 v32, v32, v36, s4
; VI-NEXT:    v_lshlrev_b32_e32 v36, 16, v41
; VI-NEXT:    v_or_b32_e32 v31, v31, v36
; VI-NEXT:    buffer_store_dword v31, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v31, v55, v61, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v31, v32, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v29, v29, v44, s4
; VI-NEXT:    v_perm_b32 v27, v27, v33, s4
; VI-NEXT:    v_readlane_b32 s87, v63, 31
; VI-NEXT:    v_readlane_b32 s86, v63, 30
; VI-NEXT:    v_readlane_b32 s85, v63, 29
; VI-NEXT:    v_readlane_b32 s84, v63, 28
; VI-NEXT:    v_readlane_b32 s83, v63, 27
; VI-NEXT:    v_readlane_b32 s82, v63, 26
; VI-NEXT:    v_readlane_b32 s81, v63, 25
; VI-NEXT:    v_readlane_b32 s80, v63, 24
; VI-NEXT:    v_readlane_b32 s71, v63, 23
; VI-NEXT:    v_readlane_b32 s70, v63, 22
; VI-NEXT:    v_readlane_b32 s69, v63, 21
; VI-NEXT:    v_readlane_b32 s68, v63, 20
; VI-NEXT:    v_readlane_b32 s67, v63, 19
; VI-NEXT:    v_readlane_b32 s66, v63, 18
; VI-NEXT:    v_readlane_b32 s65, v63, 17
; VI-NEXT:    v_readlane_b32 s64, v63, 16
; VI-NEXT:    v_readlane_b32 s55, v63, 15
; VI-NEXT:    v_readlane_b32 s54, v63, 14
; VI-NEXT:    v_readlane_b32 s53, v63, 13
; VI-NEXT:    v_readlane_b32 s52, v63, 12
; VI-NEXT:    v_readlane_b32 s51, v63, 11
; VI-NEXT:    v_readlane_b32 s50, v63, 10
; VI-NEXT:    v_readlane_b32 s49, v63, 9
; VI-NEXT:    v_readlane_b32 s48, v63, 8
; VI-NEXT:    v_readlane_b32 s39, v63, 7
; VI-NEXT:    v_readlane_b32 s38, v63, 6
; VI-NEXT:    v_readlane_b32 s37, v63, 5
; VI-NEXT:    v_readlane_b32 s36, v63, 4
; VI-NEXT:    v_readlane_b32 s35, v63, 3
; VI-NEXT:    v_readlane_b32 s34, v63, 2
; VI-NEXT:    v_readlane_b32 s31, v63, 1
; VI-NEXT:    v_readlane_b32 s30, v63, 0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v29, v29, v31
; VI-NEXT:    v_add_u32_e32 v31, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v29, v31, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v29, v30, v60, s4
; VI-NEXT:    v_perm_b32 v30, v40, v59, s4
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    v_add_u32_e32 v30, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v29, v30, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v29, v50, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v27, v27, v29
; VI-NEXT:    v_add_u32_e32 v29, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v27, v29, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v27, v28, v49, s4
; VI-NEXT:    v_perm_b32 v28, v58, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    v_add_u32_e32 v28, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v27, v28, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v28, s4
; VI-NEXT:    v_or_b32_e32 v25, v25, v27
; VI-NEXT:    v_add_u32_e32 v27, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v25, v27, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v25, v26, v35, s4
; VI-NEXT:    v_perm_b32 v26, v57, v54, s4
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    v_add_u32_e32 v26, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v25, v26, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v26, s4
; VI-NEXT:    v_or_b32_e32 v23, v23, v25
; VI-NEXT:    v_add_u32_e32 v25, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v23, v25, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v23, v24, v48, s4
; VI-NEXT:    v_perm_b32 v24, v56, v34, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    v_add_u32_e32 v24, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v23, v24, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s4
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v24, s4
; VI-NEXT:    v_or_b32_e32 v21, v21, v23
; VI-NEXT:    v_add_u32_e32 v23, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v21, v23, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v21, v22, v47, s4
; VI-NEXT:    v_perm_b32 v22, v38, v53, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    v_add_u32_e32 v22, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v21, v22, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v19, v22, s4
; VI-NEXT:    v_or_b32_e32 v19, v19, v21
; VI-NEXT:    v_add_u32_e32 v21, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v19, v21, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v19, v20, v52, s4
; VI-NEXT:    v_perm_b32 v20, v37, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    v_add_u32_e32 v20, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v19, v20, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v17, v20, s4
; VI-NEXT:    v_or_b32_e32 v17, v17, v19
; VI-NEXT:    v_add_u32_e32 v19, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v17, v19, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s4
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    v_add_u32_e32 v18, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v17, v18, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s4
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v15, v18, s4
; VI-NEXT:    v_or_b32_e32 v15, v15, v17
; VI-NEXT:    v_add_u32_e32 v17, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v15, v17, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v16, v15, s4
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s4
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v15, v16, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v16, v15, s4
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v11, v16, s4
; VI-NEXT:    v_or_b32_e32 v11, v11, v15
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v11, v15, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v12, v11, s4
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v15, v12, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v11, v12, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v9, v11, s4
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v11, v11, v15, s4
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_or_b32_e32 v9, v9, v11
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v9, v11, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v10, v9, s4
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v10, v11, v10, s4
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v9, v10, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v7, v9, s4
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v9, v9, v10, s4
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_e32 v7, v7, v9
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v7, v9, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s4
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v9, v8, s4
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v7, v8, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v5, v7, s4
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v7, v7, v8, s4
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v5, v5, v7
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v5, v7, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s4
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s4
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v3, v5, s4
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v5, v5, v6, s4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v3, v3, v5
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v3, v5, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s4
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v3, s4
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v3, v3, v4, s4
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v1, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v13, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v14, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v128i8_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v63, s30, 0
; GFX9-NEXT:    v_writelane_b32 v63, s31, 1
; GFX9-NEXT:    v_writelane_b32 v63, s34, 2
; GFX9-NEXT:    v_writelane_b32 v63, s35, 3
; GFX9-NEXT:    v_writelane_b32 v63, s36, 4
; GFX9-NEXT:    v_writelane_b32 v63, s37, 5
; GFX9-NEXT:    v_writelane_b32 v63, s38, 6
; GFX9-NEXT:    v_writelane_b32 v63, s39, 7
; GFX9-NEXT:    v_writelane_b32 v63, s48, 8
; GFX9-NEXT:    v_writelane_b32 v63, s49, 9
; GFX9-NEXT:    v_writelane_b32 v63, s50, 10
; GFX9-NEXT:    v_writelane_b32 v63, s51, 11
; GFX9-NEXT:    v_writelane_b32 v63, s52, 12
; GFX9-NEXT:    v_writelane_b32 v63, s53, 13
; GFX9-NEXT:    v_writelane_b32 v63, s54, 14
; GFX9-NEXT:    v_writelane_b32 v63, s55, 15
; GFX9-NEXT:    v_writelane_b32 v63, s64, 16
; GFX9-NEXT:    v_writelane_b32 v63, s65, 17
; GFX9-NEXT:    v_writelane_b32 v63, s66, 18
; GFX9-NEXT:    v_writelane_b32 v63, s67, 19
; GFX9-NEXT:    v_writelane_b32 v63, s68, 20
; GFX9-NEXT:    v_writelane_b32 v63, s69, 21
; GFX9-NEXT:    v_writelane_b32 v63, s70, 22
; GFX9-NEXT:    v_writelane_b32 v63, s71, 23
; GFX9-NEXT:    v_writelane_b32 v63, s80, 24
; GFX9-NEXT:    v_writelane_b32 v63, s81, 25
; GFX9-NEXT:    v_writelane_b32 v63, s82, 26
; GFX9-NEXT:    v_writelane_b32 v63, s83, 27
; GFX9-NEXT:    v_writelane_b32 v63, s84, 28
; GFX9-NEXT:    v_writelane_b32 v63, s85, 29
; GFX9-NEXT:    v_writelane_b32 v63, s86, 30
; GFX9-NEXT:    v_writelane_b32 v63, s87, 31
; GFX9-NEXT:    v_writelane_b32 v63, s96, 32
; GFX9-NEXT:    v_writelane_b32 v63, s97, 33
; GFX9-NEXT:    v_writelane_b32 v63, s98, 34
; GFX9-NEXT:    v_readfirstlane_b32 s44, v19
; GFX9-NEXT:    v_writelane_b32 v63, s99, 35
; GFX9-NEXT:    v_readfirstlane_b32 s5, v18
; GFX9-NEXT:    v_readfirstlane_b32 s4, v17
; GFX9-NEXT:    v_readfirstlane_b32 s7, v16
; GFX9-NEXT:    v_readfirstlane_b32 s6, v15
; GFX9-NEXT:    v_readfirstlane_b32 s9, v14
; GFX9-NEXT:    v_readfirstlane_b32 s8, v13
; GFX9-NEXT:    v_readfirstlane_b32 s11, v12
; GFX9-NEXT:    v_readfirstlane_b32 s10, v11
; GFX9-NEXT:    v_readfirstlane_b32 s13, v10
; GFX9-NEXT:    v_readfirstlane_b32 s12, v9
; GFX9-NEXT:    v_readfirstlane_b32 s15, v8
; GFX9-NEXT:    v_readfirstlane_b32 s14, v7
; GFX9-NEXT:    v_readfirstlane_b32 s41, v6
; GFX9-NEXT:    v_readfirstlane_b32 s40, v5
; GFX9-NEXT:    v_readfirstlane_b32 s43, v4
; GFX9-NEXT:    v_readfirstlane_b32 s42, v3
; GFX9-NEXT:    v_readfirstlane_b32 s45, v2
; GFX9-NEXT:    s_cmp_lg_u32 s44, 0
; GFX9-NEXT:    v_readfirstlane_b32 s44, v1
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; GFX9-NEXT:    s_cbranch_scc0 .LBB73_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 0
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s11, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s11, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s11, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s13, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s13, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s13, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s15, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s15, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s15, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s41, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s41, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s41, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s43, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s43, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s43, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s45, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s45, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s45, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s29, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s27, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s27, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s25, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s25, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s25, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s23, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s23, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s23, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s21, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s21, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s21, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s19, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s19, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s19, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s17, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s17, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s17, 8
; GFX9-NEXT:    s_lshr_b32 s83, s6, 16
; GFX9-NEXT:    s_lshr_b32 s82, s6, 8
; GFX9-NEXT:    s_lshr_b32 s85, s8, 16
; GFX9-NEXT:    s_lshr_b32 s84, s8, 8
; GFX9-NEXT:    s_lshr_b32 s87, s10, 16
; GFX9-NEXT:    s_lshr_b32 s86, s10, 8
; GFX9-NEXT:    s_lshr_b32 s97, s12, 16
; GFX9-NEXT:    s_lshr_b32 s96, s12, 8
; GFX9-NEXT:    s_lshr_b32 s99, s14, 16
; GFX9-NEXT:    s_lshr_b32 s98, s14, 8
; GFX9-NEXT:    s_lshr_b32 s39, s40, 16
; GFX9-NEXT:    s_lshr_b32 s38, s40, 8
; GFX9-NEXT:    s_lshr_b32 s49, s42, 16
; GFX9-NEXT:    s_lshr_b32 s48, s42, 8
; GFX9-NEXT:    s_lshr_b32 s51, s44, 16
; GFX9-NEXT:    s_lshr_b32 s50, s44, 8
; GFX9-NEXT:    s_lshr_b32 s53, s28, 16
; GFX9-NEXT:    s_lshr_b32 s52, s28, 8
; GFX9-NEXT:    s_lshr_b32 s55, s26, 16
; GFX9-NEXT:    s_lshr_b32 s54, s26, 8
; GFX9-NEXT:    s_lshr_b32 s65, s24, 16
; GFX9-NEXT:    s_lshr_b32 s64, s24, 8
; GFX9-NEXT:    s_lshr_b32 s67, s22, 16
; GFX9-NEXT:    s_lshr_b32 s66, s22, 8
; GFX9-NEXT:    s_lshr_b32 s69, s20, 16
; GFX9-NEXT:    s_lshr_b32 s68, s20, 8
; GFX9-NEXT:    s_lshr_b32 s71, s18, 16
; GFX9-NEXT:    s_lshr_b32 s70, s18, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 49
; GFX9-NEXT:    s_lshr_b32 s81, s16, 16
; GFX9-NEXT:    s_lshr_b32 s80, s16, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[16:17], 24
; GFX9-NEXT:    s_cbranch_execnz .LBB73_4
; GFX9-NEXT:  .LBB73_2: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[11:12], s[4:5], 1.0
; GFX9-NEXT:    v_add_f64 v[1:2], s[6:7], 1.0
; GFX9-NEXT:    v_add_f64 v[3:4], s[8:9], 1.0
; GFX9-NEXT:    v_add_f64 v[5:6], s[10:11], 1.0
; GFX9-NEXT:    v_add_f64 v[7:8], s[12:13], 1.0
; GFX9-NEXT:    v_add_f64 v[9:10], s[14:15], 1.0
; GFX9-NEXT:    v_add_f64 v[13:14], s[40:41], 1.0
; GFX9-NEXT:    v_add_f64 v[17:18], s[42:43], 1.0
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[11:12]
; GFX9-NEXT:    v_add_f64 v[21:22], s[44:45], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[1:2]
; GFX9-NEXT:    v_add_f64 v[23:24], s[28:29], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[3:4]
; GFX9-NEXT:    v_add_f64 v[28:29], s[26:27], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
; GFX9-NEXT:    v_add_f64 v[32:33], s[24:25], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
; GFX9-NEXT:    v_add_f64 v[34:35], s[22:23], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[9:10]
; GFX9-NEXT:    v_add_f64 v[36:37], s[20:21], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[13:14]
; GFX9-NEXT:    v_add_f64 v[38:39], s[18:19], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[17:18]
; GFX9-NEXT:    v_add_f64 v[48:49], s[16:17], 1.0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[21:22]
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 16, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[23:24]
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 8, v22
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[28:29]
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 8, v24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[32:33]
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v29
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[34:35]
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 16, v28
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[36:37]
; GFX9-NEXT:    v_lshrrev_b32_e32 v31, 16, v33
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[38:39]
; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 24, v35
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[48:49]
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 16, v35
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v3
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v3
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v5
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v5
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v7
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v7
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v9
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v9
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v18
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v18
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v18
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v17
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v17
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v22
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v22
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v21
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v21
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v23
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v23
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v29
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v29
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v28
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v33
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v32
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v32
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 8, v33
; GFX9-NEXT:    v_lshrrev_b32_e32 v30, 8, v35
; GFX9-NEXT:    v_lshrrev_b32_e32 v27, 16, v34
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v25, 24, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v41, 16, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 16, v36
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 8, v36
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v39
; GFX9-NEXT:    v_lshrrev_b32_e32 v26, 16, v39
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v39
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 16, v38
; GFX9-NEXT:    v_lshrrev_b32_e32 v40, 8, v38
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 24, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 16, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 8, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 16, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v48
; GFX9-NEXT:    s_branch .LBB73_5
; GFX9-NEXT:  .LBB73_3:
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr80
; GFX9-NEXT:    ; implicit-def: $sgpr81
; GFX9-NEXT:    ; implicit-def: $sgpr70
; GFX9-NEXT:    ; implicit-def: $sgpr71
; GFX9-NEXT:    ; implicit-def: $sgpr68
; GFX9-NEXT:    ; implicit-def: $sgpr69
; GFX9-NEXT:    ; implicit-def: $sgpr66
; GFX9-NEXT:    ; implicit-def: $sgpr67
; GFX9-NEXT:    ; implicit-def: $sgpr64
; GFX9-NEXT:    ; implicit-def: $sgpr65
; GFX9-NEXT:    ; implicit-def: $sgpr54
; GFX9-NEXT:    ; implicit-def: $sgpr55
; GFX9-NEXT:    ; implicit-def: $sgpr52
; GFX9-NEXT:    ; implicit-def: $sgpr53
; GFX9-NEXT:    ; implicit-def: $sgpr50
; GFX9-NEXT:    ; implicit-def: $sgpr51
; GFX9-NEXT:    ; implicit-def: $sgpr48
; GFX9-NEXT:    ; implicit-def: $sgpr49
; GFX9-NEXT:    ; implicit-def: $sgpr38
; GFX9-NEXT:    ; implicit-def: $sgpr39
; GFX9-NEXT:    ; implicit-def: $sgpr98
; GFX9-NEXT:    ; implicit-def: $sgpr99
; GFX9-NEXT:    ; implicit-def: $sgpr96
; GFX9-NEXT:    ; implicit-def: $sgpr97
; GFX9-NEXT:    ; implicit-def: $sgpr86
; GFX9-NEXT:    ; implicit-def: $sgpr87
; GFX9-NEXT:    ; implicit-def: $sgpr84
; GFX9-NEXT:    ; implicit-def: $sgpr85
; GFX9-NEXT:    ; implicit-def: $sgpr82
; GFX9-NEXT:    ; implicit-def: $sgpr83
; GFX9-NEXT:    ; implicit-def: $sgpr36
; GFX9-NEXT:    ; implicit-def: $sgpr34
; GFX9-NEXT:    ; implicit-def: $sgpr30
; GFX9-NEXT:    ; implicit-def: $sgpr94
; GFX9-NEXT:    ; implicit-def: $sgpr92
; GFX9-NEXT:    ; implicit-def: $sgpr90
; GFX9-NEXT:    ; implicit-def: $sgpr88
; GFX9-NEXT:    ; implicit-def: $sgpr78
; GFX9-NEXT:    ; implicit-def: $sgpr76
; GFX9-NEXT:    ; implicit-def: $sgpr74
; GFX9-NEXT:    ; implicit-def: $sgpr72
; GFX9-NEXT:    ; implicit-def: $sgpr62
; GFX9-NEXT:    ; implicit-def: $sgpr60
; GFX9-NEXT:    ; implicit-def: $sgpr58
; GFX9-NEXT:    ; implicit-def: $sgpr56
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    s_branch .LBB73_2
; GFX9-NEXT:  .LBB73_4:
; GFX9-NEXT:    v_mov_b32_e32 v15, s66
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s65
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s64
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s54
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s53
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s52
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s50
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s49
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s48
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s39
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s38
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s99
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s98
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s97
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s96
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s87
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s86
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s85
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s84
; GFX9-NEXT:    v_mov_b32_e32 v11, s4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s83
; GFX9-NEXT:    v_readlane_b32 s4, v62, 0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s82
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 3
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 5
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 7
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 9
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 15
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 16
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 17
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 18
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 19
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 20
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 21
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 22
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 23
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 25
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 26
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 27
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 28
; GFX9-NEXT:    v_mov_b32_e32 v55, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 29
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 30
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 31
; GFX9-NEXT:    v_mov_b32_e32 v58, s4
; GFX9-NEXT:    v_mov_b32_e32 v57, s34
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v57, s30
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v57, s94
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v57, s92
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v52, s36
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v57, s90
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v57, v51
; GFX9-NEXT:    v_mov_b32_e32 v51, s88
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s78
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s76
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s74
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s72
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s62
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s60
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s58
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s56
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readlane_b32 s4, v62, 32
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 33
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 34
; GFX9-NEXT:    v_mov_b32_e32 v50, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 35
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 36
; GFX9-NEXT:    v_mov_b32_e32 v31, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 37
; GFX9-NEXT:    v_mov_b32_e32 v16, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 38
; GFX9-NEXT:    v_mov_b32_e32 v54, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 39
; GFX9-NEXT:    v_mov_b32_e32 v47, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 40
; GFX9-NEXT:    v_mov_b32_e32 v30, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 41
; GFX9-NEXT:    v_mov_b32_e32 v25, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 42
; GFX9-NEXT:    v_mov_b32_e32 v41, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 43
; GFX9-NEXT:    v_mov_b32_e32 v44, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 44
; GFX9-NEXT:    v_mov_b32_e32 v42, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 45
; GFX9-NEXT:    v_mov_b32_e32 v26, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 46
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 47
; GFX9-NEXT:    v_mov_b32_e32 v61, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 48
; GFX9-NEXT:    v_mov_b32_e32 v51, s46
; GFX9-NEXT:    v_mov_b32_e32 v43, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 49
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v12, s5
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    v_mov_b32_e32 v2, s7
; GFX9-NEXT:    v_mov_b32_e32 v3, s8
; GFX9-NEXT:    v_mov_b32_e32 v4, s9
; GFX9-NEXT:    v_mov_b32_e32 v5, s10
; GFX9-NEXT:    v_mov_b32_e32 v6, s11
; GFX9-NEXT:    v_mov_b32_e32 v7, s12
; GFX9-NEXT:    v_mov_b32_e32 v8, s13
; GFX9-NEXT:    v_mov_b32_e32 v9, s14
; GFX9-NEXT:    v_mov_b32_e32 v10, s15
; GFX9-NEXT:    v_mov_b32_e32 v13, s40
; GFX9-NEXT:    v_mov_b32_e32 v14, s41
; GFX9-NEXT:    v_mov_b32_e32 v17, s42
; GFX9-NEXT:    v_mov_b32_e32 v18, s43
; GFX9-NEXT:    v_mov_b32_e32 v21, s44
; GFX9-NEXT:    v_mov_b32_e32 v22, s45
; GFX9-NEXT:    v_mov_b32_e32 v23, s28
; GFX9-NEXT:    v_mov_b32_e32 v24, s29
; GFX9-NEXT:    v_mov_b32_e32 v28, s26
; GFX9-NEXT:    v_mov_b32_e32 v29, s27
; GFX9-NEXT:    v_mov_b32_e32 v32, s24
; GFX9-NEXT:    v_mov_b32_e32 v33, s25
; GFX9-NEXT:    v_mov_b32_e32 v34, s22
; GFX9-NEXT:    v_mov_b32_e32 v35, s23
; GFX9-NEXT:    v_mov_b32_e32 v36, s20
; GFX9-NEXT:    v_mov_b32_e32 v37, s21
; GFX9-NEXT:    v_mov_b32_e32 v38, s18
; GFX9-NEXT:    v_mov_b32_e32 v39, s19
; GFX9-NEXT:    v_mov_b32_e32 v48, s16
; GFX9-NEXT:    v_mov_b32_e32 v49, s17
; GFX9-NEXT:    v_mov_b32_e32 v59, s81
; GFX9-NEXT:    v_mov_b32_e32 v60, s80
; GFX9-NEXT:    v_mov_b32_e32 v56, s71
; GFX9-NEXT:    v_mov_b32_e32 v40, s70
; GFX9-NEXT:    v_mov_b32_e32 v46, s69
; GFX9-NEXT:    v_mov_b32_e32 v19, s68
; GFX9-NEXT:    v_mov_b32_e32 v27, s67
; GFX9-NEXT:    v_mov_b32_e32 v20, s4
; GFX9-NEXT:    v_mov_b32_e32 v51, s55
; GFX9-NEXT:  .LBB73_5: ; %end
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v39, v39, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v49, v49, v20, s4
; GFX9-NEXT:    v_perm_b32 v36, v36, v19, s4
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v30, v35, v30, s4
; GFX9-NEXT:    v_perm_b32 v16, v33, v16, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v48, v48, v60, s4
; GFX9-NEXT:    v_perm_b32 v25, v41, v25, s4
; GFX9-NEXT:    v_perm_b32 v37, v37, v44, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    v_or_b32_e32 v25, v37, v25
; GFX9-NEXT:    v_perm_b32 v26, v26, v42, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX9-NEXT:    v_perm_b32 v38, v38, v40, s4
; GFX9-NEXT:    v_or_b32_e32 v26, v39, v26
; GFX9-NEXT:    v_perm_b32 v24, v24, v58, s4
; GFX9-NEXT:    v_perm_b32 v22, v22, v55, s4
; GFX9-NEXT:    v_readlane_b32 s99, v63, 35
; GFX9-NEXT:    v_readlane_b32 s98, v63, 34
; GFX9-NEXT:    v_readlane_b32 s97, v63, 33
; GFX9-NEXT:    v_readlane_b32 s96, v63, 32
; GFX9-NEXT:    v_readlane_b32 s87, v63, 31
; GFX9-NEXT:    v_readlane_b32 s86, v63, 30
; GFX9-NEXT:    v_readlane_b32 s85, v63, 29
; GFX9-NEXT:    v_readlane_b32 s84, v63, 28
; GFX9-NEXT:    v_readlane_b32 s83, v63, 27
; GFX9-NEXT:    v_readlane_b32 s82, v63, 26
; GFX9-NEXT:    v_readlane_b32 s81, v63, 25
; GFX9-NEXT:    v_readlane_b32 s80, v63, 24
; GFX9-NEXT:    v_readlane_b32 s71, v63, 23
; GFX9-NEXT:    v_readlane_b32 s70, v63, 22
; GFX9-NEXT:    v_readlane_b32 s69, v63, 21
; GFX9-NEXT:    v_readlane_b32 s68, v63, 20
; GFX9-NEXT:    v_readlane_b32 s67, v63, 19
; GFX9-NEXT:    v_readlane_b32 s66, v63, 18
; GFX9-NEXT:    v_readlane_b32 s65, v63, 17
; GFX9-NEXT:    v_readlane_b32 s64, v63, 16
; GFX9-NEXT:    v_readlane_b32 s55, v63, 15
; GFX9-NEXT:    v_readlane_b32 s54, v63, 14
; GFX9-NEXT:    v_readlane_b32 s53, v63, 13
; GFX9-NEXT:    v_readlane_b32 s52, v63, 12
; GFX9-NEXT:    v_readlane_b32 s51, v63, 11
; GFX9-NEXT:    v_readlane_b32 s50, v63, 10
; GFX9-NEXT:    v_readlane_b32 s49, v63, 9
; GFX9-NEXT:    v_readlane_b32 s48, v63, 8
; GFX9-NEXT:    v_readlane_b32 s39, v63, 7
; GFX9-NEXT:    v_readlane_b32 s38, v63, 6
; GFX9-NEXT:    v_readlane_b32 s37, v63, 5
; GFX9-NEXT:    v_readlane_b32 s36, v63, 4
; GFX9-NEXT:    v_readlane_b32 s35, v63, 3
; GFX9-NEXT:    v_readlane_b32 s34, v63, 2
; GFX9-NEXT:    v_readlane_b32 s31, v63, 1
; GFX9-NEXT:    v_readlane_b32 s30, v63, 0
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v34, v34, v15, s4
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v27, v27, v19, s4
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v45, v59, v52, s4
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v35, v15, v19, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v56, v56, v59, s4
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v19, v51, v19, s4
; GFX9-NEXT:    v_perm_b32 v32, v32, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v40, v46, v59, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v28, v15, s4
; GFX9-NEXT:    v_perm_b32 v28, v29, v50, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v15, v50, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v23, v15, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v33, v15, v50, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v21, v15, s4
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v50, v15, v50, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v17, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v15, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v15, v15, v51, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v51, 16, v45
; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v27
; GFX9-NEXT:    v_or_b32_e32 v48, v48, v51
; GFX9-NEXT:    v_or_b32_e32 v25, v34, v25
; GFX9-NEXT:    buffer_store_dword v48, v0, s[0:3], 0 offen
; GFX9-NEXT:    v_perm_b32 v48, v43, v61, s4
; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    v_perm_b32 v25, v47, v54, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    v_or_b32_e32 v48, v49, v48
; GFX9-NEXT:    v_or_b32_e32 v25, v30, v25
; GFX9-NEXT:    buffer_store_dword v48, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshlrev_b32_e32 v48, 16, v56
; GFX9-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v40
; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v35
; GFX9-NEXT:    v_or_b32_e32 v38, v38, v48
; GFX9-NEXT:    v_or_b32_e32 v26, v36, v26
; GFX9-NEXT:    v_or_b32_e32 v25, v32, v25
; GFX9-NEXT:    buffer_store_dword v38, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_store_dword v26, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v31, v25, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v25
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v19
; GFX9-NEXT:    v_or_b32_e32 v16, v20, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v19, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v28, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v29
; GFX9-NEXT:    v_or_b32_e32 v16, v23, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v19, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v24, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v33
; GFX9-NEXT:    v_or_b32_e32 v16, v21, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v19, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v22, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v50
; GFX9-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v18, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v16, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v15
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v9, v13, s4
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v13
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v13, v10, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v7, v9, s4
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v9
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s4
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v7, s4
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v7, v7, v8, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s4
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v3, v5, s4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v5, v5, v6, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v5
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v3, s4
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v3, v3, v4, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v57, v2, s4
; GFX9-NEXT:    v_perm_b32 v1, v11, v1, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v12, v1, s4
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v128i8_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:72
; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:76
; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v74, s30, 0
; GFX11-NEXT:    v_writelane_b32 v75, s96, 0
; GFX11-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-NEXT:    v_readfirstlane_b32 s41, v14
; GFX11-NEXT:    v_readfirstlane_b32 s40, v13
; GFX11-NEXT:    v_writelane_b32 v74, s31, 1
; GFX11-NEXT:    v_writelane_b32 v75, s97, 1
; GFX11-NEXT:    v_readfirstlane_b32 s15, v12
; GFX11-NEXT:    v_readfirstlane_b32 s14, v11
; GFX11-NEXT:    v_readfirstlane_b32 s11, v10
; GFX11-NEXT:    v_writelane_b32 v74, s34, 2
; GFX11-NEXT:    v_writelane_b32 v75, s98, 2
; GFX11-NEXT:    v_readfirstlane_b32 s10, v9
; GFX11-NEXT:    v_readfirstlane_b32 s9, v8
; GFX11-NEXT:    v_readfirstlane_b32 s8, v7
; GFX11-NEXT:    v_writelane_b32 v74, s35, 3
; GFX11-NEXT:    v_writelane_b32 v75, s99, 3
; GFX11-NEXT:    v_readfirstlane_b32 s7, v6
; GFX11-NEXT:    v_readfirstlane_b32 s6, v5
; GFX11-NEXT:    v_readfirstlane_b32 s5, v4
; GFX11-NEXT:    v_writelane_b32 v74, s36, 4
; GFX11-NEXT:    v_writelane_b32 v75, s100, 4
; GFX11-NEXT:    v_readfirstlane_b32 s4, v3
; GFX11-NEXT:    v_readfirstlane_b32 s13, v2
; GFX11-NEXT:    v_readfirstlane_b32 s12, v1
; GFX11-NEXT:    v_writelane_b32 v74, s37, 5
; GFX11-NEXT:    v_writelane_b32 v75, s101, 5
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:4
; GFX11-NEXT:    scratch_store_b32 off, v73, s32
; GFX11-NEXT:    v_writelane_b32 v74, s38, 6
; GFX11-NEXT:    v_writelane_b32 v75, s102, 6
; GFX11-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
; GFX11-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
; GFX11-NEXT:    v_writelane_b32 v74, s39, 7
; GFX11-NEXT:    v_writelane_b32 v75, s103, 7
; GFX11-NEXT:    v_writelane_b32 v74, s48, 8
; GFX11-NEXT:    v_writelane_b32 v75, s104, 8
; GFX11-NEXT:    v_writelane_b32 v74, s49, 9
; GFX11-NEXT:    v_writelane_b32 v74, s50, 10
; GFX11-NEXT:    v_writelane_b32 v74, s51, 11
; GFX11-NEXT:    v_writelane_b32 v74, s52, 12
; GFX11-NEXT:    v_writelane_b32 v74, s53, 13
; GFX11-NEXT:    v_writelane_b32 v74, s54, 14
; GFX11-NEXT:    v_writelane_b32 v74, s55, 15
; GFX11-NEXT:    v_writelane_b32 v74, s64, 16
; GFX11-NEXT:    v_writelane_b32 v74, s65, 17
; GFX11-NEXT:    v_writelane_b32 v74, s66, 18
; GFX11-NEXT:    v_writelane_b32 v74, s67, 19
; GFX11-NEXT:    v_writelane_b32 v74, s68, 20
; GFX11-NEXT:    v_writelane_b32 v74, s69, 21
; GFX11-NEXT:    v_writelane_b32 v74, s70, 22
; GFX11-NEXT:    v_writelane_b32 v74, s71, 23
; GFX11-NEXT:    v_writelane_b32 v74, s80, 24
; GFX11-NEXT:    v_writelane_b32 v74, s81, 25
; GFX11-NEXT:    v_writelane_b32 v74, s82, 26
; GFX11-NEXT:    v_writelane_b32 v74, s83, 27
; GFX11-NEXT:    v_writelane_b32 v74, s84, 28
; GFX11-NEXT:    v_writelane_b32 v74, s85, 29
; GFX11-NEXT:    v_writelane_b32 v74, s86, 30
; GFX11-NEXT:    v_writelane_b32 v74, s87, 31
; GFX11-NEXT:    s_cbranch_scc0 .LBB73_3
; GFX11-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-NEXT:    s_lshr_b32 s42, s11, 16
; GFX11-NEXT:    s_lshr_b32 s48, s41, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s9, 24
; GFX11-NEXT:    s_lshr_b32 s39, s41, 16
; GFX11-NEXT:    s_lshr_b32 s38, s41, 8
; GFX11-NEXT:    s_lshr_b32 s50, s40, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s9, 16
; GFX11-NEXT:    s_lshr_b32 s49, s40, 8
; GFX11-NEXT:    s_lshr_b32 s37, s15, 24
; GFX11-NEXT:    s_lshr_b32 s36, s15, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s9, 8
; GFX11-NEXT:    s_lshr_b32 s35, s15, 8
; GFX11-NEXT:    s_lshr_b32 s52, s14, 16
; GFX11-NEXT:    s_lshr_b32 s51, s14, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s7, 24
; GFX11-NEXT:    s_lshr_b32 s34, s11, 24
; GFX11-NEXT:    s_lshr_b32 s104, s11, 8
; GFX11-NEXT:    s_lshr_b32 s54, s10, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s7, 16
; GFX11-NEXT:    s_lshr_b32 s53, s10, 8
; GFX11-NEXT:    s_lshr_b32 s64, s8, 16
; GFX11-NEXT:    s_lshr_b32 s55, s8, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s7, 8
; GFX11-NEXT:    s_lshr_b32 s66, s6, 16
; GFX11-NEXT:    s_lshr_b32 s65, s6, 8
; GFX11-NEXT:    s_lshr_b32 s68, s4, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s5, 24
; GFX11-NEXT:    s_lshr_b32 s67, s4, 8
; GFX11-NEXT:    s_lshr_b32 s70, s12, 16
; GFX11-NEXT:    s_lshr_b32 s69, s12, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s5, 16
; GFX11-NEXT:    s_lshr_b32 s80, s28, 16
; GFX11-NEXT:    s_lshr_b32 s71, s28, 8
; GFX11-NEXT:    s_lshr_b32 s82, s26, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s5, 8
; GFX11-NEXT:    s_lshr_b32 s81, s26, 8
; GFX11-NEXT:    s_lshr_b32 s84, s24, 16
; GFX11-NEXT:    s_lshr_b32 s83, s24, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s13, 24
; GFX11-NEXT:    s_lshr_b32 s86, s22, 16
; GFX11-NEXT:    s_lshr_b32 s85, s22, 8
; GFX11-NEXT:    s_lshr_b32 s96, s20, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s13, 16
; GFX11-NEXT:    s_lshr_b32 s87, s20, 8
; GFX11-NEXT:    s_lshr_b32 s98, s18, 16
; GFX11-NEXT:    s_lshr_b32 s97, s18, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s13, 8
; GFX11-NEXT:    s_lshr_b32 s100, s16, 16
; GFX11-NEXT:    s_lshr_b32 s99, s16, 8
; GFX11-NEXT:    s_lshr_b32 s102, s2, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s29, 24
; GFX11-NEXT:    s_lshr_b32 s101, s2, 8
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s0, 16
; GFX11-NEXT:    s_lshr_b32 s103, s0, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s29, 16
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[40:41], 24
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[10:11], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s29, 8
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[8:9], 24
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[6:7], 24
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[4:5], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[28:29], 24
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[26:27], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s27, 8
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s25, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v77, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
; GFX11-NEXT:    v_writelane_b32 v77, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v77, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
; GFX11-NEXT:    v_writelane_b32 v77, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v77, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s1, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v77, s42, 7
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-NEXT:    s_cbranch_vccnz .LBB73_4
; GFX11-NEXT:  .LBB73_2: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[21:22], s[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[11:12], s[4:5], 1.0
; GFX11-NEXT:    v_add_f64 v[23:24], s[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[13:14], s[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[27:28], s[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[15:16], s[28:29], 1.0
; GFX11-NEXT:    v_add_f64 v[31:32], s[16:17], 1.0
; GFX11-NEXT:    v_add_f64 v[7:8], s[8:9], 1.0
; GFX11-NEXT:    v_add_f64 v[17:18], s[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[35:36], s[2:3], 1.0
; GFX11-NEXT:    v_add_f64 v[1:2], s[40:41], 1.0
; GFX11-NEXT:    v_add_f64 v[3:4], s[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[5:6], s[10:11], 1.0
; GFX11-NEXT:    v_add_f64 v[9:10], s[6:7], 1.0
; GFX11-NEXT:    v_add_f64 v[19:20], s[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[48:49], s[0:1], 1.0
; GFX11-NEXT:    v_lshrrev_b64 v[64:65], 24, v[21:22]
; GFX11-NEXT:    v_lshrrev_b64 v[50:51], 24, v[11:12]
; GFX11-NEXT:    v_lshrrev_b64 v[65:66], 24, v[23:24]
; GFX11-NEXT:    v_lshrrev_b64 v[51:52], 24, v[13:14]
; GFX11-NEXT:    v_lshrrev_b64 v[66:67], 24, v[27:28]
; GFX11-NEXT:    v_lshrrev_b64 v[52:53], 24, v[15:16]
; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[31:32]
; GFX11-NEXT:    v_lshrrev_b64 v[37:38], 24, v[7:8]
; GFX11-NEXT:    v_lshrrev_b64 v[53:54], 24, v[17:18]
; GFX11-NEXT:    v_lshrrev_b64 v[68:69], 24, v[35:36]
; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[1:2]
; GFX11-NEXT:    v_lshrrev_b64 v[29:30], 24, v[3:4]
; GFX11-NEXT:    v_lshrrev_b64 v[33:34], 24, v[5:6]
; GFX11-NEXT:    v_lshrrev_b64 v[38:39], 24, v[9:10]
; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
; GFX11-NEXT:    v_lshrrev_b64 v[69:70], 24, v[48:49]
; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 24, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 16, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v34, 8, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v30, 16, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 8, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 24, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 16, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 8, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 8, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 24, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 16, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 8, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 8, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 24, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 16, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 8, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 16, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 8, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 24, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 16, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 8, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 16, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 8, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 24, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 16, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 8, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 16, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 8, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 24, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 16, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 8, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 16, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 8, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 24, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 16, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 8, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 16, v15
; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 8, v15
; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 24, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 16, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 8, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 16, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 8, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 24, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 16, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 16, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 8, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 24, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 16, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 8, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 16, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 24, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 16, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 8, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 16, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 8, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 24, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 16, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 8, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 16, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 8, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 24, v32
; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 16, v32
; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 8, v32
; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 16, v31
; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 8, v31
; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 24, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 16, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 8, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v35
; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 8, v35
; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 24, v49
; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 16, v49
; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 8, v49
; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 16, v48
; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 8, v48
; GFX11-NEXT:    s_branch .LBB73_5
; GFX11-NEXT:  .LBB73_3:
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr103
; GFX11-NEXT:    ; implicit-def: $vcc_hi
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr101
; GFX11-NEXT:    ; implicit-def: $sgpr102
; GFX11-NEXT:    ; implicit-def: $sgpr44
; GFX11-NEXT:    ; implicit-def: $sgpr99
; GFX11-NEXT:    ; implicit-def: $sgpr100
; GFX11-NEXT:    ; implicit-def: $sgpr46
; GFX11-NEXT:    ; implicit-def: $sgpr97
; GFX11-NEXT:    ; implicit-def: $sgpr98
; GFX11-NEXT:    ; implicit-def: $sgpr56
; GFX11-NEXT:    ; implicit-def: $sgpr87
; GFX11-NEXT:    ; implicit-def: $sgpr96
; GFX11-NEXT:    ; implicit-def: $sgpr58
; GFX11-NEXT:    ; implicit-def: $sgpr85
; GFX11-NEXT:    ; implicit-def: $sgpr86
; GFX11-NEXT:    ; implicit-def: $sgpr60
; GFX11-NEXT:    ; implicit-def: $sgpr83
; GFX11-NEXT:    ; implicit-def: $sgpr84
; GFX11-NEXT:    ; implicit-def: $sgpr81
; GFX11-NEXT:    ; implicit-def: $sgpr82
; GFX11-NEXT:    ; implicit-def: $sgpr71
; GFX11-NEXT:    ; implicit-def: $sgpr80
; GFX11-NEXT:    ; implicit-def: $sgpr69
; GFX11-NEXT:    ; implicit-def: $sgpr70
; GFX11-NEXT:    ; implicit-def: $sgpr67
; GFX11-NEXT:    ; implicit-def: $sgpr68
; GFX11-NEXT:    ; implicit-def: $sgpr65
; GFX11-NEXT:    ; implicit-def: $sgpr66
; GFX11-NEXT:    ; implicit-def: $sgpr55
; GFX11-NEXT:    ; implicit-def: $sgpr64
; GFX11-NEXT:    ; implicit-def: $sgpr53
; GFX11-NEXT:    ; implicit-def: $sgpr54
; GFX11-NEXT:    ; implicit-def: $sgpr104
; GFX11-NEXT:    ; implicit-def: $sgpr34
; GFX11-NEXT:    ; implicit-def: $sgpr51
; GFX11-NEXT:    ; implicit-def: $sgpr52
; GFX11-NEXT:    ; implicit-def: $sgpr35
; GFX11-NEXT:    ; implicit-def: $sgpr36
; GFX11-NEXT:    ; implicit-def: $sgpr37
; GFX11-NEXT:    ; implicit-def: $sgpr49
; GFX11-NEXT:    ; implicit-def: $sgpr50
; GFX11-NEXT:    ; implicit-def: $sgpr38
; GFX11-NEXT:    ; implicit-def: $sgpr39
; GFX11-NEXT:    ; implicit-def: $sgpr48
; GFX11-NEXT:    ; implicit-def: $sgpr30
; GFX11-NEXT:    ; implicit-def: $sgpr94
; GFX11-NEXT:    ; implicit-def: $sgpr92
; GFX11-NEXT:    ; implicit-def: $sgpr90
; GFX11-NEXT:    ; implicit-def: $sgpr88
; GFX11-NEXT:    ; implicit-def: $sgpr78
; GFX11-NEXT:    ; implicit-def: $sgpr76
; GFX11-NEXT:    ; implicit-def: $sgpr74
; GFX11-NEXT:    ; implicit-def: $sgpr72
; GFX11-NEXT:    ; implicit-def: $sgpr62
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    s_branch .LBB73_2
; GFX11-NEXT:  .LBB73_4:
; GFX11-NEXT:    v_dual_mov_b32 v48, s0 :: v_dual_mov_b32 v49, s1
; GFX11-NEXT:    v_readlane_b32 s0, v76, 0
; GFX11-NEXT:    v_mov_b32_e32 v83, s36
; GFX11-NEXT:    v_dual_mov_b32 v1, s40 :: v_dual_mov_b32 v2, s41
; GFX11-NEXT:    v_dual_mov_b32 v3, s14 :: v_dual_mov_b32 v4, s15
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-NEXT:    v_mov_b32_e32 v98, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 1
; GFX11-NEXT:    v_mov_b32_e32 v97, s34
; GFX11-NEXT:    v_dual_mov_b32 v5, s10 :: v_dual_mov_b32 v6, s11
; GFX11-NEXT:    v_dual_mov_b32 v7, s8 :: v_dual_mov_b32 v8, s9
; GFX11-NEXT:    v_mov_b32_e32 v102, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 2
; GFX11-NEXT:    v_dual_mov_b32 v9, s6 :: v_dual_mov_b32 v10, s7
; GFX11-NEXT:    v_dual_mov_b32 v11, s4 :: v_dual_mov_b32 v12, s5
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v103, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 3
; GFX11-NEXT:    v_dual_mov_b32 v13, s12 :: v_dual_mov_b32 v14, s13
; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
; GFX11-NEXT:    v_mov_b32_e32 v101, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 4
; GFX11-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
; GFX11-NEXT:    v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v114, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 5
; GFX11-NEXT:    v_dual_mov_b32 v21, s22 :: v_dual_mov_b32 v22, s23
; GFX11-NEXT:    v_dual_mov_b32 v23, s20 :: v_dual_mov_b32 v24, s21
; GFX11-NEXT:    v_mov_b32_e32 v116, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 6
; GFX11-NEXT:    v_dual_mov_b32 v27, s18 :: v_dual_mov_b32 v28, s19
; GFX11-NEXT:    v_dual_mov_b32 v31, s16 :: v_dual_mov_b32 v32, s17
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v115, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 7
; GFX11-NEXT:    v_dual_mov_b32 v35, s2 :: v_dual_mov_b32 v36, s3
; GFX11-NEXT:    v_dual_mov_b32 v41, vcc_hi :: v_dual_mov_b32 v42, s103
; GFX11-NEXT:    v_mov_b32_e32 v130, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 8
; GFX11-NEXT:    v_dual_mov_b32 v181, s102 :: v_dual_mov_b32 v180, s101
; GFX11-NEXT:    v_dual_mov_b32 v178, s100 :: v_dual_mov_b32 v177, s99
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v131, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 9
; GFX11-NEXT:    v_dual_mov_b32 v166, s98 :: v_dual_mov_b32 v165, s97
; GFX11-NEXT:    v_dual_mov_b32 v161, s96 :: v_dual_mov_b32 v160, s87
; GFX11-NEXT:    v_mov_b32_e32 v129, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 10
; GFX11-NEXT:    v_dual_mov_b32 v151, s86 :: v_dual_mov_b32 v146, s84
; GFX11-NEXT:    v_dual_mov_b32 v149, s85 :: v_dual_mov_b32 v132, s81
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v134, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 11
; GFX11-NEXT:    v_dual_mov_b32 v135, s83 :: v_dual_mov_b32 v128, s80
; GFX11-NEXT:    v_dual_mov_b32 v133, s82 :: v_dual_mov_b32 v118, s70
; GFX11-NEXT:    v_mov_b32_e32 v145, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 12
; GFX11-NEXT:    v_dual_mov_b32 v119, s71 :: v_dual_mov_b32 v112, s67
; GFX11-NEXT:    v_dual_mov_b32 v117, s69 :: v_dual_mov_b32 v100, s66
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v144, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 13
; GFX11-NEXT:    v_dual_mov_b32 v113, s68 :: v_dual_mov_b32 v96, s64
; GFX11-NEXT:    v_dual_mov_b32 v99, s65 :: v_dual_mov_b32 v84, s53
; GFX11-NEXT:    v_mov_b32_e32 v148, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 14
; GFX11-NEXT:    v_dual_mov_b32 v87, s55 :: v_dual_mov_b32 v80, s52
; GFX11-NEXT:    v_dual_mov_b32 v85, s54 :: v_dual_mov_b32 v30, s50
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v150, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 15
; GFX11-NEXT:    v_dual_mov_b32 v71, s51 :: v_dual_mov_b32 v26, s49
; GFX11-NEXT:    v_dual_mov_b32 v39, s48 :: v_dual_mov_b32 v34, s38
; GFX11-NEXT:    v_mov_b32_e32 v147, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 16
; GFX11-NEXT:    v_dual_mov_b32 v55, s39 :: v_dual_mov_b32 v82, s37
; GFX11-NEXT:    v_dual_mov_b32 v81, s35 :: v_dual_mov_b32 v86, s104
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v162, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 17
; GFX11-NEXT:    v_dual_mov_b32 v69, s42 :: v_dual_mov_b32 v68, s44
; GFX11-NEXT:    v_dual_mov_b32 v67, s46 :: v_dual_mov_b32 v66, s56
; GFX11-NEXT:    v_mov_b32_e32 v164, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 18
; GFX11-NEXT:    v_dual_mov_b32 v65, s58 :: v_dual_mov_b32 v64, s60
; GFX11-NEXT:    v_dual_mov_b32 v54, s30 :: v_dual_mov_b32 v53, s94
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v163, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 19
; GFX11-NEXT:    v_dual_mov_b32 v52, s92 :: v_dual_mov_b32 v51, s90
; GFX11-NEXT:    v_dual_mov_b32 v50, s88 :: v_dual_mov_b32 v37, s76
; GFX11-NEXT:    v_mov_b32_e32 v176, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 20
; GFX11-NEXT:    v_dual_mov_b32 v38, s78 :: v_dual_mov_b32 v33, s74
; GFX11-NEXT:    v_mov_b32_e32 v29, s72
; GFX11-NEXT:    v_mov_b32_e32 v25, s62
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v179, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 21
; GFX11-NEXT:    v_mov_b32_e32 v167, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 22
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v182, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 23
; GFX11-NEXT:    v_mov_b32_e32 v40, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 24
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v183, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 25
; GFX11-NEXT:    v_mov_b32_e32 v44, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 26
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v45, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 27
; GFX11-NEXT:    v_mov_b32_e32 v43, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 28
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v46, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 29
; GFX11-NEXT:    v_mov_b32_e32 v56, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 30
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v47, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 31
; GFX11-NEXT:    v_mov_b32_e32 v58, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v59, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 1
; GFX11-NEXT:    v_mov_b32_e32 v57, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v60, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 3
; GFX11-NEXT:    v_mov_b32_e32 v62, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v61, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 5
; GFX11-NEXT:    v_mov_b32_e32 v63, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v72, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 7
; GFX11-NEXT:    v_mov_b32_e32 v73, s0
; GFX11-NEXT:  .LBB73_5: ; %end
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-NEXT:    v_perm_b32 v69, v41, v69, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v48, v48, v42, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v70, v72, v63, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v68, v181, v68, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v181, v62, v60, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
; GFX11-NEXT:    v_perm_b32 v49, v49, v73, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 16, v70
; GFX11-NEXT:    v_perm_b32 v35, v35, v180, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v36, v36, v61, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v180, 16, v181
; GFX11-NEXT:    v_or_b32_e32 v60, v48, v69
; GFX11-NEXT:    v_perm_b32 v48, v178, v67, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v61, v49, v70
; GFX11-NEXT:    v_perm_b32 v49, v56, v46, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v63, v36, v180
; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 16, v68
; GFX11-NEXT:    v_lshlrev_b32_e32 v36, 16, v48
; GFX11-NEXT:    v_perm_b32 v48, v166, v66, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v31, v31, v177, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v27, v27, v165, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v28, v28, v47, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; GFX11-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; GFX11-NEXT:    v_or_b32_e32 v62, v35, v68
; GFX11-NEXT:    v_perm_b32 v35, v59, v58, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v66, v31, v36
; GFX11-NEXT:    v_or_b32_e32 v69, v28, v49
; GFX11-NEXT:    v_or_b32_e32 v68, v27, v48
; GFX11-NEXT:    v_perm_b32 v27, v161, v65, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v28, v45, v44, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v31, v151, v64, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v32, v32, v57, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-NEXT:    v_perm_b32 v23, v23, v160, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-NEXT:    v_perm_b32 v24, v24, v43, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-NEXT:    v_perm_b32 v21, v21, v149, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-NEXT:    v_or_b32_e32 v67, v32, v35
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_store_b128 v0, v[60:63], off
; GFX11-NEXT:    scratch_store_b128 v0, v[66:69], off offset:16
; GFX11-NEXT:    v_or_b32_e32 v64, v23, v27
; GFX11-NEXT:    v_perm_b32 v23, v40, v182, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v65, v24, v28
; GFX11-NEXT:    v_or_b32_e32 v66, v21, v31
; GFX11-NEXT:    v_perm_b32 v21, v22, v183, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v22, v146, v54, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v179, v176, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v27, v133, v53, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v19, v19, v135, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v20, v20, v167, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v17, v17, v132, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-NEXT:    v_or_b32_e32 v67, v21, v23
; GFX11-NEXT:    v_or_b32_e32 v19, v19, v22
; GFX11-NEXT:    v_perm_b32 v22, v164, v162, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v20, v20, v24
; GFX11-NEXT:    v_or_b32_e32 v21, v17, v27
; GFX11-NEXT:    v_perm_b32 v17, v18, v163, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v18, v128, v52, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v118, v51, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v150, v148, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v15, v15, v119, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v13, v13, v117, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v16, v16, v147, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_or_b32_e32 v22, v17, v22
; GFX11-NEXT:    v_or_b32_e32 v15, v15, v18
; GFX11-NEXT:    v_perm_b32 v18, v145, v134, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v17, v13, v24
; GFX11-NEXT:    v_perm_b32 v24, v100, v38, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v16, v16, v23
; GFX11-NEXT:    v_perm_b32 v13, v14, v144, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v14, v113, v50, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v131, v130, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v9, v9, v99, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v11, v11, v112, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v12, v12, v129, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_or_b32_e32 v18, v13, v18
; GFX11-NEXT:    v_or_b32_e32 v13, v9, v24
; GFX11-NEXT:    v_perm_b32 v9, v10, v115, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v10, v96, v37, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v11, v11, v14
; GFX11-NEXT:    v_perm_b32 v14, v116, v114, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v12, v12, v23
; GFX11-NEXT:    v_perm_b32 v23, v103, v102, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v85, v33, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v7, v7, v87, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v8, v8, v101, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v5, v5, v84, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_or_b32_e32 v7, v7, v10
; GFX11-NEXT:    v_perm_b32 v10, v98, v97, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v14, v9, v14
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v23
; GFX11-NEXT:    v_or_b32_e32 v9, v5, v24
; GFX11-NEXT:    v_perm_b32 v5, v80, v29, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v83, v82, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v30, v25, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v25, v55, v39, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v6, v6, v86, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_perm_b32 v3, v3, v71, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-NEXT:    v_perm_b32 v4, v4, v81, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v26, v1, v26, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v27, v2, v34, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-NEXT:    v_or_b32_e32 v10, v6, v10
; GFX11-NEXT:    v_or_b32_e32 v1, v3, v5
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v23
; GFX11-NEXT:    v_or_b32_e32 v3, v26, v24
; GFX11-NEXT:    v_or_b32_e32 v4, v27, v25
; GFX11-NEXT:    s_clause 0x5
; GFX11-NEXT:    scratch_store_b128 v0, v[64:67], off offset:32
; GFX11-NEXT:    scratch_store_b128 v0, v[19:22], off offset:48
; GFX11-NEXT:    scratch_store_b128 v0, v[15:18], off offset:64
; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v73, off, s32
; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-NEXT:    v_readlane_b32 s104, v75, 8
; GFX11-NEXT:    v_readlane_b32 s103, v75, 7
; GFX11-NEXT:    v_readlane_b32 s102, v75, 6
; GFX11-NEXT:    v_readlane_b32 s101, v75, 5
; GFX11-NEXT:    v_readlane_b32 s100, v75, 4
; GFX11-NEXT:    v_readlane_b32 s99, v75, 3
; GFX11-NEXT:    v_readlane_b32 s98, v75, 2
; GFX11-NEXT:    v_readlane_b32 s97, v75, 1
; GFX11-NEXT:    v_readlane_b32 s96, v75, 0
; GFX11-NEXT:    v_readlane_b32 s87, v74, 31
; GFX11-NEXT:    v_readlane_b32 s86, v74, 30
; GFX11-NEXT:    v_readlane_b32 s85, v74, 29
; GFX11-NEXT:    v_readlane_b32 s84, v74, 28
; GFX11-NEXT:    v_readlane_b32 s83, v74, 27
; GFX11-NEXT:    v_readlane_b32 s82, v74, 26
; GFX11-NEXT:    v_readlane_b32 s81, v74, 25
; GFX11-NEXT:    v_readlane_b32 s80, v74, 24
; GFX11-NEXT:    v_readlane_b32 s71, v74, 23
; GFX11-NEXT:    v_readlane_b32 s70, v74, 22
; GFX11-NEXT:    v_readlane_b32 s69, v74, 21
; GFX11-NEXT:    v_readlane_b32 s68, v74, 20
; GFX11-NEXT:    v_readlane_b32 s67, v74, 19
; GFX11-NEXT:    v_readlane_b32 s66, v74, 18
; GFX11-NEXT:    v_readlane_b32 s65, v74, 17
; GFX11-NEXT:    v_readlane_b32 s64, v74, 16
; GFX11-NEXT:    v_readlane_b32 s55, v74, 15
; GFX11-NEXT:    v_readlane_b32 s54, v74, 14
; GFX11-NEXT:    v_readlane_b32 s53, v74, 13
; GFX11-NEXT:    v_readlane_b32 s52, v74, 12
; GFX11-NEXT:    v_readlane_b32 s51, v74, 11
; GFX11-NEXT:    v_readlane_b32 s50, v74, 10
; GFX11-NEXT:    v_readlane_b32 s49, v74, 9
; GFX11-NEXT:    v_readlane_b32 s48, v74, 8
; GFX11-NEXT:    v_readlane_b32 s39, v74, 7
; GFX11-NEXT:    v_readlane_b32 s38, v74, 6
; GFX11-NEXT:    v_readlane_b32 s37, v74, 5
; GFX11-NEXT:    v_readlane_b32 s36, v74, 4
; GFX11-NEXT:    v_readlane_b32 s35, v74, 3
; GFX11-NEXT:    v_readlane_b32 s34, v74, 2
; GFX11-NEXT:    v_readlane_b32 s31, v74, 1
; GFX11-NEXT:    v_readlane_b32 s30, v74, 0
; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:72
; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:76
; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) {
; SI-LABEL: bitcast_v128i8_to_v16f64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:388
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:184
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v1
; SI-NEXT:    v_lshlrev_b32_e32 v42, 24, v3
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v5
; SI-NEXT:    v_lshlrev_b32_e32 v40, 24, v7
; SI-NEXT:    v_lshlrev_b32_e32 v53, 8, v9
; SI-NEXT:    v_lshlrev_b32_e32 v34, 24, v11
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v13
; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v15
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v17
; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v19
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v21
; SI-NEXT:    v_lshlrev_b32_e32 v58, 24, v23
; SI-NEXT:    v_lshlrev_b32_e32 v56, 8, v25
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v27
; SI-NEXT:    v_lshlrev_b32_e32 v46, 8, v29
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v30
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v0
; SI-NEXT:    v_lshlrev_b32_e32 v44, 8, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v33
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v35
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v49
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v50
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v51
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v52
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:188
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:52
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:196
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:204
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:212
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:220
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:228
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:260
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:268
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:292
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:316
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:332
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:352
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:376
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:348
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:356
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:384
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380
; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v3
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:364
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:372
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB74_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v49
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    v_or_b32_e32 v2, v2, v41
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_or_b32_e32 v3, v40, v3
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v2, v2, v53
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v48
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v62
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v38
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v56
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v46
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v35
; SI-NEXT:    v_or_b32_e32 v8, v8, v44
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v33
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v52
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v63
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v39
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v61
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v51
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v50
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v45
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v36
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v43
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:  .LBB74_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB74_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_movk_i32 s6, 0x300
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v49
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b32 s7, 0x3000000
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v54, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v40, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v32, v4
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v60, v5
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v58, v6
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v37, v7
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v41, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v55, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v53, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_or_b32_e32 v3, v48, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v62, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v38, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v56, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v7, v46, v7
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v35
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v8, v44, v8
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v33
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v9
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v39
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v59
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v51
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v45
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v18
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v21
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v22
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v27
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
; SI-NEXT:    v_or_b32_e32 v31, v43, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x300, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v31
; SI-NEXT:  .LBB74_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v16f64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:12
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:380
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:124
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB74_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s6, 0xc0c0004
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v9, v33, v62, s6
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_perm_b32 v10, v63, v60, s6
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v11, v53, v58, s6
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_perm_b32 v12, v51, v56, s6
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v37, v36, s6
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v45, v44, s6
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v49, v48, s6
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v0, v1, v0, s6
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s6
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v41, s6
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_perm_b32 v8, v55, v34, s6
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_perm_b32 v9, v61, v32, s6
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_perm_b32 v10, v59, v54, s6
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_perm_b32 v11, v57, v52, s6
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v47, v38, s6
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_perm_b32 v13, v35, v46, s6
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_perm_b32 v14, v43, v50, s6
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_perm_b32 v15, v39, v42, s6
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v30, v30, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v32, v40, v32, s6
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:  .LBB74_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB74_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v31, 0x300
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v41
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v62
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v60
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v58
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v56
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v36
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v44
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v48
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v0, 0x300, v0
; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    s_waitcnt vmcnt(11)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    s_waitcnt vmcnt(10)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    s_waitcnt vmcnt(9)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    s_waitcnt vmcnt(8)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    s_waitcnt vmcnt(7)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    s_waitcnt vmcnt(6)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v1, v1, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v3, v3, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v4, v4, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v5, v5, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v6, v6, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v7, v7, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v8, v8, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_add_u16_e32 v8, 3, v55
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v34
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 3, v33
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
; VI-NEXT:    v_add_u16_sdwa v9, v9, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_add_u16_e32 v9, 3, v61
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v32
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 3, v63
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
; VI-NEXT:    v_add_u16_sdwa v10, v10, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_add_u16_e32 v10, 3, v59
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v54
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 3, v53
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
; VI-NEXT:    v_add_u16_sdwa v11, v11, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_add_u16_e32 v11, 3, v57
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v52
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 3, v51
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
; VI-NEXT:    v_add_u16_sdwa v12, v12, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_add_u16_e32 v12, 3, v47
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 3, v37
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
; VI-NEXT:    v_add_u16_sdwa v13, v13, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_add_u16_e32 v13, 3, v35
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v46
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 3, v45
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
; VI-NEXT:    v_add_u16_sdwa v14, v14, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_add_u16_e32 v14, 3, v43
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v50
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v15, 3, v49
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
; VI-NEXT:    v_add_u16_sdwa v15, v15, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    v_add_u16_e32 v15, 3, v39
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v42
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v16, v16, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v18, v18, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v17, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v18, 0x300, v18
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v19, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v20, v20, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v20, 0x300, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v21, v21, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v22, v22, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v23, v23, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v24, v24, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v25, v25, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v26, v26, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v27, v27, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v28, v28, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v29, v29, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v30, v30, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    v_or_b32_sdwa v30, v32, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v32, v32, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v30, v32
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; VI-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v33, 3, v40
; VI-NEXT:    v_or_b32_sdwa v33, v34, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v32, 0x300, v32
; VI-NEXT:    v_add_u16_sdwa v31, v33, v31 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:  .LBB74_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v16f64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:12
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(32)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:380
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:376
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:128
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB74_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0xc0c0004
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v10, v55, v32, s6
; GFX9-NEXT:    v_perm_b32 v11, v61, v60, s6
; GFX9-NEXT:    v_perm_b32 v12, v53, v38, s6
; GFX9-NEXT:    v_perm_b32 v13, v57, v56, s6
; GFX9-NEXT:    v_perm_b32 v14, v51, v34, s6
; GFX9-NEXT:    v_perm_b32 v15, v45, v50, s6
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v16, v43, v16, s6
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v9, s6
; GFX9-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX9-NEXT:    v_perm_b32 v8, v35, v42, s6
; GFX9-NEXT:    v_perm_b32 v9, v41, v48, s6
; GFX9-NEXT:    v_lshl_or_b32 v8, v9, 16, v8
; GFX9-NEXT:    v_perm_b32 v9, v39, v40, s6
; GFX9-NEXT:    v_lshl_or_b32 v9, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v10, v62, v63, s6
; GFX9-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX9-NEXT:    v_perm_b32 v11, v59, v54, s6
; GFX9-NEXT:    v_lshl_or_b32 v11, v12, 16, v11
; GFX9-NEXT:    v_perm_b32 v12, v58, v37, s6
; GFX9-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v47, v52, s6
; GFX9-NEXT:    v_lshl_or_b32 v13, v14, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v33, v46, s6
; GFX9-NEXT:    v_lshl_or_b32 v14, v15, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v49, v44, s6
; GFX9-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s6
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    v_lshl_or_b32 v16, v17, 16, v16
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v18, v17, s6
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    v_lshl_or_b32 v18, v19, 16, v18
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    v_lshl_or_b32 v19, v20, 16, v19
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    v_lshl_or_b32 v21, v22, 16, v21
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    v_lshl_or_b32 v23, v24, 16, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    v_lshl_or_b32 v24, v25, 16, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s6
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    v_lshl_or_b32 v25, v26, 16, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    v_lshl_or_b32 v26, v27, 16, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s6
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    v_lshl_or_b32 v28, v29, 16, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s6
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    v_lshl_or_b32 v29, v30, 16, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    v_lshl_or_b32 v30, v31, 16, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v32, v36, v32, s6
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    v_lshl_or_b32 v31, v32, 16, v31
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:  .LBB74_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB74_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    s_movk_i32 s6, 0x300
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v48
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v32
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v60
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v38
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v56
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v34
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v50
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(25)
; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(24)
; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v0
; GFX9-NEXT:    s_waitcnt vmcnt(22)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(19)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(14)
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(13)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    s_waitcnt vmcnt(11)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    s_waitcnt vmcnt(8)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_sdwa v1, v1, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v3, v3, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v4, v4, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v5, v5, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v6, v6, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v7, v7, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v8, v8, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v35
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v42
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v41
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u16_sdwa v9, v9, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v39
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v40
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v55
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u16_sdwa v10, v10, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v62
; GFX9-NEXT:    v_lshlrev_b16_e32 v11, 8, v63
; GFX9-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v61
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u16_sdwa v11, v11, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v59
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v54
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v53
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u16_sdwa v12, v12, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v58
; GFX9-NEXT:    v_lshlrev_b16_e32 v13, 8, v37
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v57
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u16_sdwa v13, v13, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v12, v12, v13
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v47
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v52
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v51
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    v_add_u16_sdwa v14, v14, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v33
; GFX9-NEXT:    v_lshlrev_b16_e32 v15, 8, v46
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v45
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    v_add_u16_sdwa v15, v15, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v14, v14, v15
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v49
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v44
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v43
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_add_u16_sdwa v16, v16, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v15, v15, v16
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v33, 8, v33
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v17, v17, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v17
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v18
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v18, v18, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v19, v19, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v18, v18, v19
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v20, v20, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v19, v19, v20
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v21, v21, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v20, v20, v21
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v22, v22, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v21, v21, v22
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v23, v23, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v24, 8, v24
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v24, v24, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v23, v23, v24
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v25, 8, v25
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v25, v25, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v24, v24, v25
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v26, 8, v26
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v26, v26, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v27, v27, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v26, v26, v27
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v28, v28, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v29, 8, v29
; GFX9-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v29, v29, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v28, v28, v29
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v30, 8, v30
; GFX9-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v30, v30, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v29, v29, v30
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v31, 8, v31
; GFX9-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_sdwa v31, v31, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; GFX9-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v32, 3, v36
; GFX9-NEXT:    v_or_b32_sdwa v32, v33, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v31
; GFX9-NEXT:    v_add_u16_sdwa v32, v32, s6 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_e32 v31, v31, v32
; GFX9-NEXT:  .LBB74_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v16f64:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:384
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:380
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:376
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:372
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:368
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:364
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:360
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:356
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:352
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:348
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:344
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:340
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:336
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:332
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:328
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:260
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:132
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v40, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v42, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v45, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v47, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v57, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v58, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:388
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v60, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:8
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v74.l, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v73.l, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v75.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v76.l, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v77.l, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v78.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v88.l, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v79.l, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v89.l, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v90.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v92.l, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v91.l, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v93.l, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v94.l, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v95.l, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v104.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v106.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v105.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v107.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v108.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v110.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v109.l, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v111.l, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v120.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v121.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v122.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v123.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v124.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v125.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v126.l, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v127.l, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB74_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v127, v126, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v125, v124, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v123, v122, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v121, v120, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v107, v105, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v106, v104, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v95, v94, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v111, v109, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v110, v108, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v93, v91, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v92, v90, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v89, v79, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v88, v78, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v77, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v75, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v74, v72, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v63, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v62, v60, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v59, v58, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v57, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v56, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v45, v43, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v44, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v41, v40, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v183, v181, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v182, v180, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v179, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v178, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v164, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v161, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v160, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v149, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v147, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v146, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v135, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v134, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v131, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v129, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v128, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v117, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v116, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v113, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v103, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v102, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v99, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v98, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v87, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v84, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v81, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v80, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v69, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v67, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v66, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v55, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v54, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v48, v49, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v37, v39, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v38, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v36, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v29, v48, 16, v31
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v30, v35, 16, v37
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v31, v32, 16, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr127_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr126_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr125_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr124_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr122_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr121_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr120_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr111_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr109_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr108_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr107_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr105_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr106_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr104_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr95_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr94_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr93_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr91_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr92_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr90_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr89_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr88_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT:  .LBB74_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB74_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v127.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v125.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v123.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v0.h, 8, v126.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v124.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v122.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v121.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v0.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v120.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v111.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v110.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v109.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v3.h, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v108.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, v107.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, v106.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v95.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v2.l, v3.l, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v2.h, v3.h, v2.h
; GFX11-TRUE16-NEXT:    v_and_b16 v3.l, 0xff, v4.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v105.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v4.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v104.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, v92.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v90.l
; GFX11-TRUE16-NEXT:    v_or_b16 v3.l, v3.h, v3.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, v89.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v3.h, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.l, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v4.h, 8, v94.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, v93.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, v88.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v77.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.l, v4.h, v4.l
; GFX11-TRUE16-NEXT:    v_and_b16 v4.h, 0xff, v5.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v5.l, 8, v91.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, v74.l, 3
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v72.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, v63.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, v62.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v4.h, v5.l, v4.h
; GFX11-TRUE16-NEXT:    v_or_b16 v5.l, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b16 v5.h, 0xff, v6.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v6.l, 8, v79.l
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v78.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v59.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v5.h, v6.l, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, v56.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.l, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v6.h, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.l, 8, v76.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, v75.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v10.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v46.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, v45.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v6.h, v7.l, v6.h
; GFX11-TRUE16-NEXT:    v_and_b16 v7.l, 0xff, v7.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v7.h, 8, v73.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, v44.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v41.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, v182.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v180.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.l, v7.h, v7.l
; GFX11-TRUE16-NEXT:    v_or_b16 v7.h, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v9.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v8.h, 8, v61.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v9.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v60.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, v179.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.h, v8.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, v178.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v9.h, 8, v58.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, v57.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v167.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, v164.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v162.l
; GFX11-TRUE16-NEXT:    v_or_b16 v9.l, v9.h, v9.l
; GFX11-TRUE16-NEXT:    v_and_b16 v9.h, 0xff, v10.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v10.l, 8, v47.l
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, v161.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, v160.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v149.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v9.h, v10.l, v9.h
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v11.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v11.l, 8, v43.l
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v42.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, v146.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v144.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, v135.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.l, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.l, 8, v40.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, v183.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, v134.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v131.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v11.h, v12.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b16 v12.l, 0xff, v12.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v12.h, 8, v181.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, v128.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v118.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, v117.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, v116.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.h, v12.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v14.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v13.h, 8, v177.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v14.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v176.l
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v113.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.l, v13.h, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, v102.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v13.h, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.l, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v14.h, 8, v166.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, v165.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v23.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v100.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, v99.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v14.h, v14.l
; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v15.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v15.l, 8, v163.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, v98.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v87.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, v84.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v82.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v15.l, v14.h
; GFX11-TRUE16-NEXT:    v_or_b16 v15.l, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v16.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v16.l, 8, v151.l
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v150.l
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, v81.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v15.h, v16.l, v15.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, v80.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v16.h, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.l, 8, v148.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, v147.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v69.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, v66.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v28.h, 8, v64.l
; GFX11-TRUE16-NEXT:    v_or_b16 v16.h, v17.l, v16.h
; GFX11-TRUE16-NEXT:    v_and_b16 v17.l, 0xff, v17.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v17.h, 8, v145.l
; GFX11-TRUE16-NEXT:    v_and_b16 v28.l, 0xff, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, v55.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v29.h, 8, v53.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.l, 8, v51.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.l, v17.h, v17.l
; GFX11-TRUE16-NEXT:    v_or_b16 v17.h, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v18.l, 0xff, v19.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v18.h, 8, v133.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v19.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v132.l
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.h, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, v52.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v18.l, v18.h, v18.l
; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v18.h, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.l, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v19.h, 8, v130.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, v129.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v28.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v50.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.l, v29.h, v29.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.l, v19.h, v19.l
; GFX11-TRUE16-NEXT:    v_and_b16 v19.h, 0xff, v20.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v20.l, 8, v119.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, v38.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, v36.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v33.h, v34.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.h, 8, v37.l
; GFX11-TRUE16-NEXT:    v_or_b16 v19.h, v20.l, v19.h
; GFX11-TRUE16-NEXT:    v_or_b16 v20.l, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v20.h, 0xff, v21.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v21.l, 8, v115.l
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v114.l
; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v31.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.l, 8, v35.l
; GFX11-TRUE16-NEXT:    v_or_b16 v20.h, v21.l, v20.h
; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v21.l, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v21.h, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.l, 8, v112.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, v103.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.l, 8, v33.l
; GFX11-TRUE16-NEXT:    v_and_b16 v33.h, 0xff, v33.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.l, 8, v32.l
; GFX11-TRUE16-NEXT:    v_or_b16 v21.h, v22.l, v21.h
; GFX11-TRUE16-NEXT:    v_and_b16 v22.l, 0xff, v22.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v22.h, 8, v101.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.l, v34.l, v31.l
; GFX11-TRUE16-NEXT:    v_or_b16 v31.h, v33.l, v31.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v32.l, v33.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.l, v22.h, v22.l
; GFX11-TRUE16-NEXT:    v_or_b16 v22.h, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_and_b16 v23.l, 0xff, v24.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v23.h, 8, v97.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v24.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v96.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v23.l, v23.h, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v23.h, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v24.h, 8, v86.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, v85.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.h, v24.l
; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v25.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v25.l, 8, v83.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v3.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v4.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v4.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v5.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v25.l, v24.h
; GFX11-TRUE16-NEXT:    v_or_b16 v25.l, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_and_b16 v25.h, 0xff, v26.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v26.l, 8, v71.l
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v70.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v5.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v6.l
; GFX11-TRUE16-NEXT:    v_or_b16 v25.h, v26.l, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v6.h
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.l, 8, v68.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v67.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v7.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v7.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v8.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v27.l, v26.h
; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v27.h, 8, v65.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v8.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v9.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v9.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v27.l, v27.h, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, v54.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v10.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v11.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v11.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v12.l
; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v27.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v12.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v13.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v13.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v14.l
; GFX11-TRUE16-NEXT:    v_or_b16 v29.h, v30.l, v27.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v28.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, v49.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v28.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v29.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, v39.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v30.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v30.h, 8, v48.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v14.h
; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v15.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v15.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v30.h, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v16.l
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v32.h, v29.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v16.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v17.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v17.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v18.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v19.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v19.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v20.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v20.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v21.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v21.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v22.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v22.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v23.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v23.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v24.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v24.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v25.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v25.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v26.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v26.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v27.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v30.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v30.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v31.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v31.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v32.l
; GFX11-TRUE16-NEXT:  .LBB74_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v16f64:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v86, v0
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:384
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:380
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:376
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:372
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:368
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:364
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:360
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:356
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:352
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:348
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:344
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:340
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:336
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:332
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:328
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:260
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:188
; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v45, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v46, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v47, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v57, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v56, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v59, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v61, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v62, off, s32 offset:132
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v63, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v73, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v72, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v74, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v75, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v79, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v89, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v88, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v90, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v91, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v93, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v94, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v95, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v105, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_u16 v104, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v106, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v107, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v109, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v108, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v110, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v111, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v120, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_u16 v121, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v122, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v0, off, s32 offset:388
; GFX11-FAKE16-NEXT:    scratch_load_u16 v123, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v125, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v124, off, s32 offset:8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_u16 v126, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v127, off, s32
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, v30 :: v_dual_mov_b32 v32, v29
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v28 :: v_dual_mov_b32 v35, v27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v37, v26 :: v_dual_mov_b32 v36, v25
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v38, v24 :: v_dual_mov_b32 v39, v23
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v49, v22 :: v_dual_mov_b32 v48, v21
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v50, v20 :: v_dual_mov_b32 v51, v19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v18 :: v_dual_mov_b32 v53, v17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v16 :: v_dual_mov_b32 v55, v15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v65, v14 :: v_dual_mov_b32 v64, v13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, v12 :: v_dual_mov_b32 v67, v11
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, v10 :: v_dual_mov_b32 v68, v9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, v8 :: v_dual_mov_b32 v71, v7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, v6 :: v_dual_mov_b32 v81, v5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, v4 :: v_dual_mov_b32 v83, v3
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, v2 :: v_dual_mov_b32 v84, v1
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(5)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB74_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v86, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v85, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v82, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v80, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v66, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v65, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v54, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v69, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v52, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v50, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v49, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v38, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v37, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v7, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v9, 16, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v33, v127, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v126, v124, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v125, v123, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v122, v121, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v120, v111, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v110, v108, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v109, v107, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v106, v104, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v105, v95, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v94, v92, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v93, v91, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v90, v88, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v89, v79, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v78, v77, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v76, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v74, v72, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v73, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v62, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v61, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v58, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v57, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v46, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v45, v43, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v42, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v40, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v182, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v181, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v178, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v177, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v26, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v166, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v165, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v162, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v161, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v150, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v148, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v146, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v145, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v134, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v133, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v29, 16, v28
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v31, 16, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v130, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v129, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v118, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v117, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v115, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v112, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v103, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v34, v101, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v99, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v97, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v30, 16, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v32, 16, v31
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v34, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v36, 16, v35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr127
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr126
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr124
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr125
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr122
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr121
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr120
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr111
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr108
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr109
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr107
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr106
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr105
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr95
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr93
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr91
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr89
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:  .LBB74_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB74_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v86, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v85, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v84
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v82, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v80, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v81
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v69, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v70, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v67
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v68
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v66, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v54, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v65, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v55
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v7, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v52, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v7, 8, v51
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v50, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v49, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v7, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v8, 8, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, v38, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, v37, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v9, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v9, 8, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v34, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, v33, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v10, 8, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v7
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v127
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v126, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, v125, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v11, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v11, 8, v124
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v123
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v122, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v12, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v121
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v120, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v12, 8, v111
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v110, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, v109, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v107
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v12, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v13, 8, v108
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, v106, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, v105, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v14, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v14, 8, v104
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v95
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 16, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v10, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v94, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v12, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, v93, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v15, 8, v92
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xff, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v91
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v90, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, v89, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v16, 8, v88
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v79
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v78, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v17, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v77
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v76, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v14, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v17, 8, v75
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v74, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, v73, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v63
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v18, 8, v72
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, v62, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, v61, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v19, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v19, 8, v60
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v59
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v58, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, v57, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v20, 8, v56
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v47
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v46, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, v45, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v21, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v21, 8, v44
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v43
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v42, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v22, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v41
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v40, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v22, 8, v183
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v182, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, v181, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v179
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v23, 8, v180
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, v178, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, v177, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v24, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v24, 8, v176
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v167
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v166, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v22, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, v165, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v25, 8, v164
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v163
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v162, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v21, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, v161, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v26, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v26, 8, v160
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v151
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v150, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v27, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v149
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v148, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v24, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v27, 8, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v146, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, v145, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v135
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v28, 8, v144
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v134, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v133, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v29, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v29, 8, v132
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v131
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, v130, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, v129, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v30, 8, v128
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v31, 8, v119
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v26, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, v118, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, v117, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v31, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, v115, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v116
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v33, 8, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v113
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v32, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, v112, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, v103, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, v101, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, v99, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v38, v97, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v34, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xff, v32
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v102
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v37, 8, v100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v39, 8, v98
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v48, 8, v96
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v49, 8, v87
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v34, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v37, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v39, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v48, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v49, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v33, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v35, v36
; GFX11-FAKE16-NEXT:  .LBB74_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define inreg <16 x double> @bitcast_v128i8_to_v16f64_scalar(<128 x i8> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v128i8_to_v16f64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s76, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    v_readfirstlane_b32 s46, v30
; SI-NEXT:    v_readfirstlane_b32 s6, v29
; SI-NEXT:    v_readfirstlane_b32 s7, v28
; SI-NEXT:    v_readfirstlane_b32 s8, v27
; SI-NEXT:    v_readfirstlane_b32 s9, v26
; SI-NEXT:    v_readfirstlane_b32 s10, v25
; SI-NEXT:    v_readfirstlane_b32 s11, v24
; SI-NEXT:    v_readfirstlane_b32 s12, v23
; SI-NEXT:    v_readfirstlane_b32 s13, v22
; SI-NEXT:    v_readfirstlane_b32 s14, v21
; SI-NEXT:    v_readfirstlane_b32 s15, v20
; SI-NEXT:    v_readfirstlane_b32 s40, v19
; SI-NEXT:    v_readfirstlane_b32 s41, v18
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:96
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:104
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:328
; SI-NEXT:    v_readfirstlane_b32 s75, v1
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:332
; SI-NEXT:    v_readfirstlane_b32 s43, v16
; SI-NEXT:    v_readfirstlane_b32 s45, v14
; SI-NEXT:    v_readfirstlane_b32 s73, v3
; SI-NEXT:    v_readfirstlane_b32 s74, v2
; SI-NEXT:    v_readfirstlane_b32 s57, v11
; SI-NEXT:    v_readfirstlane_b32 s72, v4
; SI-NEXT:    v_readfirstlane_b32 s44, v15
; SI-NEXT:    v_readfirstlane_b32 s47, v13
; SI-NEXT:    v_readfirstlane_b32 s56, v12
; SI-NEXT:    v_readfirstlane_b32 s58, v10
; SI-NEXT:    v_readfirstlane_b32 s59, v9
; SI-NEXT:    v_readfirstlane_b32 s60, v8
; SI-NEXT:    v_readfirstlane_b32 s61, v7
; SI-NEXT:    v_readfirstlane_b32 s62, v6
; SI-NEXT:    v_readfirstlane_b32 s63, v5
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v50
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v51
; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v52
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v53
; SI-NEXT:    v_lshlrev_b32_e32 v13, 24, v54
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v55
; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v40
; SI-NEXT:    v_lshlrev_b32_e32 v60, 8, v41
; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v42
; SI-NEXT:    v_lshlrev_b32_e32 v41, 8, v43
; SI-NEXT:    v_lshlrev_b32_e32 v16, 24, v46
; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v56
; SI-NEXT:    v_lshlrev_b32_e32 v55, 24, v59
; SI-NEXT:    v_lshlrev_b32_e32 v54, 8, v61
; SI-NEXT:    v_lshlrev_b32_e32 v53, 24, v44
; SI-NEXT:    v_lshlrev_b32_e32 v52, 8, v45
; SI-NEXT:    v_lshlrev_b32_e32 v51, 24, v47
; SI-NEXT:    v_lshlrev_b32_e32 v50, 8, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:316
; SI-NEXT:    v_lshlrev_b32_e32 v62, 8, v30
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 24, v29
; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v28
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v27
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 8, v26
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v25
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v45, 8, v24
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v23
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v22
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v20
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v19
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v18
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(9) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:268
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:260
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:252
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:300
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:292
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 8, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 8, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 8, v34
; SI-NEXT:    v_lshlrev_b32_e32 v34, 8, v32
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:228
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:220
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204
; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 24, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 24, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 24, v33
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:196
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188
; SI-NEXT:    v_lshlrev_b32_e32 v33, 24, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:116
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB75_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v31
; SI-NEXT:    v_or_b32_e32 v0, s4, v0
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v12
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v11, v3
; SI-NEXT:    v_or_b32_e32 v11, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v30
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v10
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v13, v1
; SI-NEXT:    v_or_b32_e32 v13, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v9
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v8
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v14, v1
; SI-NEXT:    v_or_b32_e32 v14, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v42
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v7
; SI-NEXT:    v_or_b32_e32 v0, v0, v60
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v15, v1
; SI-NEXT:    v_or_b32_e32 v15, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v6
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v63
; SI-NEXT:    v_or_b32_e32 v0, v0, v41
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v16, v1
; SI-NEXT:    v_or_b32_e32 v16, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v61
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v17
; SI-NEXT:    v_or_b32_e32 v0, v0, v40
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v55, v1
; SI-NEXT:    v_or_b32_e32 v17, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v44
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v54
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v53, v1
; SI-NEXT:    v_or_b32_e32 v18, v0, v1
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v51, v1
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v12, v2, v3
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s18, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s77, s19, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s77, s5
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s77, s21, 8
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s22, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s77, s78, s77
; SI-NEXT:    s_or_b32 s5, s5, s77
; SI-NEXT:    s_and_b32 s77, s24, 0xff
; SI-NEXT:    s_lshl_b32 s78, s25, 8
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s26, 0xff
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s79, s27, 24
; SI-NEXT:    s_and_b32 s77, s77, 0xffff
; SI-NEXT:    s_or_b32 s78, s79, s78
; SI-NEXT:    s_or_b32 s77, s77, s78
; SI-NEXT:    s_and_b32 s78, s28, 0xff
; SI-NEXT:    s_lshl_b32 s79, s29, 8
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s76, 0xff
; SI-NEXT:    s_lshl_b32 s79, s79, 16
; SI-NEXT:    s_lshl_b32 s88, s75, 24
; SI-NEXT:    s_and_b32 s78, s78, 0xffff
; SI-NEXT:    s_or_b32 s79, s88, s79
; SI-NEXT:    s_or_b32 s78, s78, s79
; SI-NEXT:    s_and_b32 s79, s74, 0xff
; SI-NEXT:    s_lshl_b32 s88, s73, 8
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s72, 0xff
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s89, s63, 24
; SI-NEXT:    s_and_b32 s79, s79, 0xffff
; SI-NEXT:    s_or_b32 s88, s89, s88
; SI-NEXT:    s_or_b32 s79, s79, s88
; SI-NEXT:    s_and_b32 s88, s62, 0xff
; SI-NEXT:    s_lshl_b32 s89, s61, 8
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s60, 0xff
; SI-NEXT:    s_lshl_b32 s89, s89, 16
; SI-NEXT:    s_lshl_b32 s90, s59, 24
; SI-NEXT:    s_and_b32 s88, s88, 0xffff
; SI-NEXT:    s_or_b32 s89, s90, s89
; SI-NEXT:    s_or_b32 s88, s88, s89
; SI-NEXT:    s_and_b32 s89, s58, 0xff
; SI-NEXT:    s_lshl_b32 s90, s57, 8
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s56, 0xff
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_lshl_b32 s91, s47, 24
; SI-NEXT:    s_and_b32 s89, s89, 0xffff
; SI-NEXT:    s_or_b32 s90, s91, s90
; SI-NEXT:    s_or_b32 s89, s89, s90
; SI-NEXT:    s_and_b32 s90, s45, 0xff
; SI-NEXT:    s_lshl_b32 s91, s44, 8
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s43, 0xff
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v54
; SI-NEXT:    v_or_b32_e32 v0, v0, v52
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v19, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v49, v1
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    s_lshl_b32 s91, s91, 16
; SI-NEXT:    s_lshl_b32 s92, s42, 24
; SI-NEXT:    s_and_b32 s90, s90, 0xffff
; SI-NEXT:    s_or_b32 s91, s92, s91
; SI-NEXT:    s_or_b32 s90, s90, s91
; SI-NEXT:    s_and_b32 s91, s41, 0xff
; SI-NEXT:    s_lshl_b32 s92, s40, 8
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s15, 0xff
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s93, s14, 24
; SI-NEXT:    s_and_b32 s91, s91, 0xffff
; SI-NEXT:    s_or_b32 s92, s93, s92
; SI-NEXT:    s_or_b32 s91, s91, s92
; SI-NEXT:    s_and_b32 s92, s13, 0xff
; SI-NEXT:    s_lshl_b32 s93, s12, 8
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s11, 0xff
; SI-NEXT:    s_lshl_b32 s93, s93, 16
; SI-NEXT:    s_lshl_b32 s94, s10, 24
; SI-NEXT:    s_and_b32 s92, s92, 0xffff
; SI-NEXT:    s_or_b32 s93, s94, s93
; SI-NEXT:    s_or_b32 s92, s92, s93
; SI-NEXT:    s_and_b32 s93, s9, 0xff
; SI-NEXT:    s_lshl_b32 s94, s8, 8
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    s_and_b32 s94, s7, 0xff
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_lshl_b32 s95, s6, 24
; SI-NEXT:    s_and_b32 s93, s93, 0xffff
; SI-NEXT:    s_or_b32 s94, s95, s94
; SI-NEXT:    s_or_b32 s93, s93, s94
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v3, s78
; SI-NEXT:    v_mov_b32_e32 v4, s79
; SI-NEXT:    v_mov_b32_e32 v5, s88
; SI-NEXT:    v_mov_b32_e32 v6, s89
; SI-NEXT:    v_mov_b32_e32 v7, s90
; SI-NEXT:    v_mov_b32_e32 v8, s91
; SI-NEXT:    v_mov_b32_e32 v9, s92
; SI-NEXT:    v_mov_b32_e32 v10, s93
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v52
; SI-NEXT:    v_or_b32_e32 v0, v0, v50
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v20, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v39, v1
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v48
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v21, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v37, v1
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v48
; SI-NEXT:    v_or_b32_e32 v0, v0, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v22, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v23
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v35, v1
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v38
; SI-NEXT:    v_or_b32_e32 v0, v0, v36
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v23, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v33, v1
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v36
; SI-NEXT:    v_or_b32_e32 v0, v0, v34
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v24, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v25
; SI-NEXT:    v_or_b32_e32 v0, v0, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v29, v1
; SI-NEXT:    v_or_b32_e32 v25, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v28, v1
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v32
; SI-NEXT:    v_or_b32_e32 v0, v0, v59
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v26, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v58
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v56
; SI-NEXT:    v_or_b32_e32 v0, v0, v57
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v27, v1
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    v_or_b32_e32 v27, v0, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v0, v0, v45
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_or_b32_e32 v28, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v46
; SI-NEXT:    v_mov_b32_e32 v45, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v60, v32
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v29, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v30, v0, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v31, v0, v1
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s77
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_branch .LBB75_3
; SI-NEXT:  .LBB75_2:
; SI-NEXT:    v_mov_b32_e32 v55, v54
; SI-NEXT:    v_mov_b32_e32 v53, v52
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_mov_b32_e32 v49, v48
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v35, v34
; SI-NEXT:    v_mov_b32_e32 v33, v32
; SI-NEXT:    v_mov_b32_e32 v62, v29
; SI-NEXT:    v_mov_b32_e32 v59, v28
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    v_mov_b32_e32 v47, v43
; SI-NEXT:    v_mov_b32_e32 v57, v27
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB75_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v43, v58
; SI-NEXT:    v_mov_b32_e32 v58, v33
; SI-NEXT:    v_mov_b32_e32 v33, v37
; SI-NEXT:    v_mov_b32_e32 v34, v39
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_cbranch_vccnz .LBB75_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_and_b32 s4, s46, 0xff
; SI-NEXT:    v_mov_b32_e32 v32, v35
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_and_b32 s17, s22, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s76, s76, 3
; SI-NEXT:    s_add_i32 s74, s74, 3
; SI-NEXT:    s_add_i32 s72, s72, 3
; SI-NEXT:    s_add_i32 s62, s62, 3
; SI-NEXT:    s_add_i32 s60, s60, 3
; SI-NEXT:    s_add_i32 s58, s58, 3
; SI-NEXT:    s_add_i32 s56, s56, 3
; SI-NEXT:    s_and_b32 s22, s56, 0xff
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_addk_i32 s12, 0x300
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    s_add_i32 s10, s10, 0x3000000
; SI-NEXT:    s_add_i32 s6, s6, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v9, s10
; SI-NEXT:    v_mov_b32_e32 v10, s6
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, s4, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, 0x3000000, v0
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s16, s18, 0xff
; SI-NEXT:    s_addk_i32 s4, 0x300
; SI-NEXT:    s_lshl_b32 s5, s19, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s5, s16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s20, 0xff
; SI-NEXT:    s_lshl_b32 s16, s21, 8
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_addk_i32 s5, 0x300
; SI-NEXT:    s_lshl_b32 s16, s23, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_or_b32 s5, s16, s5
; SI-NEXT:    s_and_b32 s16, s24, 0xff
; SI-NEXT:    s_lshl_b32 s17, s25, 8
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s18, s26, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s27, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s28, 0xff
; SI-NEXT:    s_lshl_b32 s18, s29, 8
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s19, s76, 0xff
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_lshl_b32 s18, s75, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s18, s74, 0xff
; SI-NEXT:    s_lshl_b32 s19, s73, 8
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s20, s72, 0xff
; SI-NEXT:    s_addk_i32 s18, 0x300
; SI-NEXT:    s_lshl_b32 s19, s63, 24
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_or_b32 s19, s19, s20
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_and_b32 s19, s62, 0xff
; SI-NEXT:    s_lshl_b32 s20, s61, 8
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s21, s60, 0xff
; SI-NEXT:    s_addk_i32 s19, 0x300
; SI-NEXT:    s_lshl_b32 s20, s59, 24
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_or_b32 s20, s20, s21
; SI-NEXT:    s_or_b32 s19, s20, s19
; SI-NEXT:    s_and_b32 s20, s58, 0xff
; SI-NEXT:    s_lshl_b32 s21, s57, 8
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_addk_i32 s20, 0x300
; SI-NEXT:    s_lshl_b32 s21, s47, 24
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_and_b32 s21, s45, 0xff
; SI-NEXT:    s_lshl_b32 s22, s44, 8
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    s_and_b32 s23, s43, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s22, s42, 24
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s23
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s22, s41, 0xff
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x300, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_lshl_b32 s23, s40, 8
; SI-NEXT:    s_or_b32 s22, s23, s22
; SI-NEXT:    s_addk_i32 s22, 0x300
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_or_b32 s14, s14, s22
; SI-NEXT:    s_add_i32 s4, s4, 0x3000000
; SI-NEXT:    s_add_i32 s5, s5, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 0x3000000
; SI-NEXT:    s_add_i32 s17, s17, 0x3000000
; SI-NEXT:    s_add_i32 s18, s18, 0x3000000
; SI-NEXT:    s_add_i32 s19, s19, 0x3000000
; SI-NEXT:    s_add_i32 s20, s20, 0x3000000
; SI-NEXT:    s_add_i32 s21, s21, 0x3000000
; SI-NEXT:    s_add_i32 s14, s14, 0x3000000
; SI-NEXT:    v_mov_b32_e32 v4, s18
; SI-NEXT:    v_mov_b32_e32 v5, s19
; SI-NEXT:    v_mov_b32_e32 v6, s20
; SI-NEXT:    v_mov_b32_e32 v7, s21
; SI-NEXT:    v_mov_b32_e32 v8, s14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_add_i32_e32 v12, vcc, 0x3000000, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v39, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_mov_b32_e32 v3, s17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v13, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v37, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v14, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v42, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v15, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v41, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v44
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v55, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v54
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v53, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v52
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v51, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v50
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v48
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v34, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v38
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v33, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v36
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v32, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v58
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v62, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, 0x3000000, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v60
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v59, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v26, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v57, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v27, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v1, v45, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v28, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x3000000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x300, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x3000000, v0
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:  .LBB75_5: ; %end
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v16f64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s45, v15
; VI-NEXT:    v_readfirstlane_b32 s46, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s76, v0
; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_readfirstlane_b32 s6, v30
; VI-NEXT:    v_readfirstlane_b32 s7, v29
; VI-NEXT:    v_readfirstlane_b32 s8, v28
; VI-NEXT:    v_readfirstlane_b32 s9, v27
; VI-NEXT:    v_readfirstlane_b32 s10, v26
; VI-NEXT:    v_readfirstlane_b32 s11, v25
; VI-NEXT:    v_readfirstlane_b32 s12, v24
; VI-NEXT:    v_readfirstlane_b32 s13, v23
; VI-NEXT:    v_readfirstlane_b32 s14, v22
; VI-NEXT:    v_readfirstlane_b32 s15, v21
; VI-NEXT:    v_readfirstlane_b32 s40, v20
; VI-NEXT:    v_readfirstlane_b32 s41, v19
; VI-NEXT:    v_readfirstlane_b32 s42, v18
; VI-NEXT:    v_readfirstlane_b32 s43, v17
; VI-NEXT:    v_readfirstlane_b32 s44, v16
; VI-NEXT:    v_readfirstlane_b32 s56, v12
; VI-NEXT:    v_readfirstlane_b32 s57, v11
; VI-NEXT:    v_readfirstlane_b32 s58, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s60, v8
; VI-NEXT:    v_readfirstlane_b32 s61, v7
; VI-NEXT:    v_readfirstlane_b32 s62, v6
; VI-NEXT:    v_readfirstlane_b32 s63, v5
; VI-NEXT:    v_readfirstlane_b32 s72, v4
; VI-NEXT:    v_readfirstlane_b32 s73, v3
; VI-NEXT:    v_readfirstlane_b32 s74, v2
; VI-NEXT:    v_readfirstlane_b32 s75, v1
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; VI-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; VI-NEXT:    s_and_b64 s[4:5], vcc, exec
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    s_cbranch_scc0 .LBB75_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_mov_b32_e32 v3, s23
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_mov_b32_e32 v2, s21
; VI-NEXT:    v_perm_b32 v3, s22, v3, v11
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_perm_b32 v2, s20, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_or_b32_e32 v10, v10, v12
; VI-NEXT:    v_perm_b32 v12, v14, v13, s4
; VI-NEXT:    v_perm_b32 v11, s6, v15, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v52, v51, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v54, v53, s4
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    v_perm_b32 v13, v48, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v50, v49, s4
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_perm_b32 v14, v36, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v38, v37, s4
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_perm_b32 v15, v32, v16, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v16, v34, v33, s4
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    v_perm_b32 v16, v60, v17, s4
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_perm_b32 v17, v62, v61, s4
; VI-NEXT:    v_or_b32_e32 v16, v17, v16
; VI-NEXT:    v_perm_b32 v17, v59, v18, s4
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_perm_b32 v18, v41, v40, s4
; VI-NEXT:    v_or_b32_e32 v17, v18, v17
; VI-NEXT:    v_perm_b32 v18, v57, v19, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v19, v42, v58, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_perm_b32 v19, v45, v20, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_perm_b32 v20, v56, v47, s4
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_perm_b32 v20, v44, v21, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_perm_b32 v21, v46, v26, s4
; VI-NEXT:    v_or_b32_e32 v20, v21, v20
; VI-NEXT:    v_perm_b32 v21, v31, v22, s4
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_perm_b32 v22, v24, v43, s4
; VI-NEXT:    v_or_b32_e32 v21, v22, v21
; VI-NEXT:    v_perm_b32 v22, v28, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_perm_b32 v23, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v22, v23, v22
; VI-NEXT:    v_perm_b32 v23, v25, v63, s4
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_perm_b32 v24, v27, v55, s4
; VI-NEXT:    v_or_b32_e32 v23, v24, v23
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    v_or_b32_e32 v24, v25, v24
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    v_or_b32_e32 v25, v26, v25
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    v_or_b32_e32 v26, v27, v26
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    v_or_b32_e32 v27, v28, v27
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    v_or_b32_e32 v28, v29, v28
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    v_or_b32_e32 v29, v30, v29
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    v_or_b32_e32 v30, v31, v30
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v43, v31, s4
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v43, v46, v43, s4
; VI-NEXT:    v_or_b32_e32 v31, v43, v31
; VI-NEXT:    s_mov_b64 s[4:5], 0
; VI-NEXT:    s_branch .LBB75_3
; VI-NEXT:  .LBB75_2:
; VI-NEXT:    v_mov_b32_e32 v57, v56
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_mov_b32_e32 v45, v26
; VI-NEXT:    v_mov_b32_e32 v44, v24
; VI-NEXT:    s_mov_b64 s[4:5], -1
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:  .LBB75_3: ; %Flow
; VI-NEXT:    v_mov_b32_e32 v43, v44
; VI-NEXT:    v_mov_b32_e32 v44, v45
; VI-NEXT:    v_mov_b32_e32 v45, v47
; VI-NEXT:    v_mov_b32_e32 v46, v57
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; VI-NEXT:    s_cbranch_vccnz .LBB75_5
; VI-NEXT:  ; %bb.4: ; %cmp.true
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    v_mov_b32_e32 v1, s21
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    v_mov_b32_e32 v2, s23
; VI-NEXT:    v_perm_b32 v1, s20, v1, v11
; VI-NEXT:    v_perm_b32 v2, s22, v2, v11
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x300, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x300, v2
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    s_add_i32 s76, s76, 3
; VI-NEXT:    v_mov_b32_e32 v4, s75
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_perm_b32 v4, s76, v4, v11
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x300, v3
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s74, s74, 3
; VI-NEXT:    v_mov_b32_e32 v4, s73
; VI-NEXT:    s_add_i32 s72, s72, 3
; VI-NEXT:    v_mov_b32_e32 v5, s63
; VI-NEXT:    v_perm_b32 v4, s74, v4, v11
; VI-NEXT:    v_perm_b32 v5, s72, v5, v11
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x300, v4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s62, s62, 3
; VI-NEXT:    v_mov_b32_e32 v5, s61
; VI-NEXT:    s_add_i32 s60, s60, 3
; VI-NEXT:    v_mov_b32_e32 v6, s59
; VI-NEXT:    v_perm_b32 v5, s62, v5, v11
; VI-NEXT:    v_perm_b32 v6, s60, v6, v11
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x300, v5
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s58, s58, 3
; VI-NEXT:    v_mov_b32_e32 v6, s57
; VI-NEXT:    s_add_i32 s56, s56, 3
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_perm_b32 v7, s56, v7, v11
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x300, v6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    v_mov_b32_e32 v7, s45
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    v_mov_b32_e32 v8, s43
; VI-NEXT:    v_perm_b32 v7, s46, v7, v11
; VI-NEXT:    v_perm_b32 v8, s44, v8, v11
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x300, v7
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    v_mov_b32_e32 v8, s41
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_perm_b32 v8, s42, v8, v11
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x300, v8
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    v_mov_b32_e32 v9, s13
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    v_mov_b32_e32 v10, s11
; VI-NEXT:    v_perm_b32 v9, s14, v9, v11
; VI-NEXT:    v_perm_b32 v10, s12, v10, v11
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x300, v9
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    v_mov_b32_e32 v10, s9
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    v_mov_b32_e32 v12, s7
; VI-NEXT:    v_perm_b32 v10, s10, v10, v11
; VI-NEXT:    v_perm_b32 v12, s8, v12, v11
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x300, v10
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v10, v12, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v48
; VI-NEXT:    v_perm_b32 v14, v14, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v36
; VI-NEXT:    v_perm_b32 v15, v15, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x3000000, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x3000000, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x3000000, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v6
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v7
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x3000000, v8
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x3000000, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v10
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v11, s6, v12, v11
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x300, v11
; VI-NEXT:    v_perm_b32 v16, v16, v17, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v60
; VI-NEXT:    v_perm_b32 v17, v17, v18, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v57
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_perm_b32 v18, v18, v19, s4
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v12
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v19
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v12, v12, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v12, vcc, 3, v54
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v52
; VI-NEXT:    v_perm_b32 v12, v12, v53, s4
; VI-NEXT:    v_perm_b32 v13, v13, v51, s4
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x300, v12
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v13, vcc, 3, v50
; VI-NEXT:    v_perm_b32 v13, v13, v49, s4
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x300, v13
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v14, vcc, 3, v38
; VI-NEXT:    v_perm_b32 v14, v14, v37, s4
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x300, v14
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v15, vcc, 3, v34
; VI-NEXT:    v_perm_b32 v15, v15, v33, s4
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x300, v15
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v16, vcc, 3, v62
; VI-NEXT:    v_perm_b32 v16, v16, v61, s4
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x300, v16
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v17, vcc, 3, v41
; VI-NEXT:    v_perm_b32 v17, v17, v40, s4
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x300, v17
; VI-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v18, vcc, 3, v42
; VI-NEXT:    v_perm_b32 v18, v18, v47, s4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v18
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v19, vcc, 3, v46
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x3000000, v11
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x3000000, v12
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x3000000, v13
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x3000000, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x3000000, v15
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x3000000, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x3000000, v17
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x3000000, v18
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v19, v19, v20, s4
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x300, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v20
; VI-NEXT:    v_perm_b32 v20, v20, v21, s4
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v20, vcc, 3, v45
; VI-NEXT:    v_perm_b32 v20, v20, v44, s4
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x300, v20
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x3000000, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v21
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v21, vcc, 3, v43
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v22, s4
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x300, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x3000000, v21
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v22, vcc, 3, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v22, v23, s4
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x300, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x3000000, v22
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v23, vcc, 3, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v24, s4
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x300, v23
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x3000000, v23
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v24, vcc, 3, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v24, v25, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x300, v24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x3000000, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v25, vcc, 3, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v26, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x300, v25
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x3000000, v25
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v26, vcc, 3, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v26, v27, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x300, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x3000000, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v27, vcc, 3, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v27, v28, s4
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_sdwa v27, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x3000000, v27
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v28, vcc, 3, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x300, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_sdwa v28, v29, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x3000000, v28
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v29, vcc, 3, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v29, v30, s4
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x300, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_sdwa v29, v30, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x3000000, v29
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v30, vcc, 3, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v30, v31, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_sdwa v30, v31, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x3000000, v30
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v31, vcc, 3, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v31, v32, s4
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x300, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v32, vcc, 3, v32
; VI-NEXT:    v_perm_b32 v32, v32, v33, s4
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    v_or_b32_sdwa v31, v32, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x3000000, v31
; VI-NEXT:  .LBB75_5: ; %end
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v16f64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s45, v15
; GFX9-NEXT:    v_readfirstlane_b32 s46, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s76, v0
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readfirstlane_b32 s6, v30
; GFX9-NEXT:    v_readfirstlane_b32 s7, v29
; GFX9-NEXT:    v_readfirstlane_b32 s8, v28
; GFX9-NEXT:    v_readfirstlane_b32 s9, v27
; GFX9-NEXT:    v_readfirstlane_b32 s10, v26
; GFX9-NEXT:    v_readfirstlane_b32 s11, v25
; GFX9-NEXT:    v_readfirstlane_b32 s12, v24
; GFX9-NEXT:    v_readfirstlane_b32 s13, v23
; GFX9-NEXT:    v_readfirstlane_b32 s14, v22
; GFX9-NEXT:    v_readfirstlane_b32 s15, v21
; GFX9-NEXT:    v_readfirstlane_b32 s40, v20
; GFX9-NEXT:    v_readfirstlane_b32 s41, v19
; GFX9-NEXT:    v_readfirstlane_b32 s42, v18
; GFX9-NEXT:    v_readfirstlane_b32 s43, v17
; GFX9-NEXT:    v_readfirstlane_b32 s44, v16
; GFX9-NEXT:    v_readfirstlane_b32 s56, v12
; GFX9-NEXT:    v_readfirstlane_b32 s57, v11
; GFX9-NEXT:    v_readfirstlane_b32 s58, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s60, v8
; GFX9-NEXT:    v_readfirstlane_b32 s61, v7
; GFX9-NEXT:    v_readfirstlane_b32 s62, v6
; GFX9-NEXT:    v_readfirstlane_b32 s63, v5
; GFX9-NEXT:    v_readfirstlane_b32 s72, v4
; GFX9-NEXT:    v_readfirstlane_b32 s73, v3
; GFX9-NEXT:    v_readfirstlane_b32 s74, v2
; GFX9-NEXT:    v_readfirstlane_b32 s75, v1
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:200
; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:196
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:192
; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:188
; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:184
; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:180
; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:176
; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:172
; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:168
; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:164
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:132
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:12
; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(47)
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(44)
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(46)
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(45)
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(43)
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    s_cbranch_scc0 .LBB75_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX9-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_or_b32_e32 v10, v10, v12
; GFX9-NEXT:    v_perm_b32 v12, v14, v13, s4
; GFX9-NEXT:    v_perm_b32 v11, s6, v15, v11
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    v_perm_b32 v12, v52, v51, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_perm_b32 v13, v54, v53, s4
; GFX9-NEXT:    v_or_b32_e32 v12, v13, v12
; GFX9-NEXT:    v_perm_b32 v13, v48, v39, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_perm_b32 v14, v50, v49, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v14, v13
; GFX9-NEXT:    v_perm_b32 v14, v36, v35, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_perm_b32 v15, v38, v37, s4
; GFX9-NEXT:    v_or_b32_e32 v14, v15, v14
; GFX9-NEXT:    v_perm_b32 v15, v32, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    v_perm_b32 v16, v34, v33, s4
; GFX9-NEXT:    v_or_b32_e32 v15, v16, v15
; GFX9-NEXT:    v_perm_b32 v16, v60, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_perm_b32 v17, v62, v61, s4
; GFX9-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX9-NEXT:    v_perm_b32 v17, v59, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_perm_b32 v18, v41, v40, s4
; GFX9-NEXT:    v_or_b32_e32 v17, v18, v17
; GFX9-NEXT:    v_perm_b32 v18, v57, v19, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_perm_b32 v19, v42, v58, s4
; GFX9-NEXT:    v_or_b32_e32 v18, v19, v18
; GFX9-NEXT:    v_perm_b32 v19, v45, v20, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX9-NEXT:    v_perm_b32 v20, v56, v47, s4
; GFX9-NEXT:    v_or_b32_e32 v19, v20, v19
; GFX9-NEXT:    v_perm_b32 v20, v44, v21, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX9-NEXT:    v_perm_b32 v21, v46, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v20, v21, v20
; GFX9-NEXT:    v_perm_b32 v21, v31, v22, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX9-NEXT:    v_perm_b32 v22, v24, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v21, v22, v21
; GFX9-NEXT:    v_perm_b32 v22, v28, v23, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX9-NEXT:    v_perm_b32 v23, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v22, v23, v22
; GFX9-NEXT:    v_perm_b32 v23, v25, v63, s4
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX9-NEXT:    v_perm_b32 v24, v27, v55, s4
; GFX9-NEXT:    v_or_b32_e32 v23, v24, v23
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s4
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    v_or_b32_e32 v24, v25, v24
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v26, v25, s4
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    v_or_b32_e32 v25, v26, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s4
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    v_or_b32_e32 v26, v27, v26
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v28, v27, s4
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    v_or_b32_e32 v27, v28, v27
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    v_or_b32_e32 v28, v29, v28
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v30, v29, s4
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    v_or_b32_e32 v29, v30, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    v_or_b32_e32 v30, v31, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v43, v31, s4
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v43, v46, v43, s4
; GFX9-NEXT:    v_or_b32_e32 v31, v43, v31
; GFX9-NEXT:    s_mov_b64 s[4:5], 0
; GFX9-NEXT:    s_branch .LBB75_3
; GFX9-NEXT:  .LBB75_2:
; GFX9-NEXT:    v_mov_b32_e32 v57, v56
; GFX9-NEXT:    v_mov_b32_e32 v47, v46
; GFX9-NEXT:    v_mov_b32_e32 v45, v26
; GFX9-NEXT:    v_mov_b32_e32 v44, v24
; GFX9-NEXT:    s_mov_b64 s[4:5], -1
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:  .LBB75_3: ; %Flow
; GFX9-NEXT:    v_mov_b32_e32 v43, v44
; GFX9-NEXT:    v_mov_b32_e32 v44, v45
; GFX9-NEXT:    v_mov_b32_e32 v45, v47
; GFX9-NEXT:    v_mov_b32_e32 v46, v57
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_vccnz .LBB75_5
; GFX9-NEXT:  ; %bb.4: ; %cmp.true
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v11
; GFX9-NEXT:    s_movk_i32 s4, 0x300
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v11
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    v_add_u32_sdwa v1, v1, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v11
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v11
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v2
; GFX9-NEXT:    v_add_u32_sdwa v3, v3, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v2, v2, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    s_add_i32 s76, s76, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s75
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v11
; GFX9-NEXT:    v_perm_b32 v4, s76, v4, v11
; GFX9-NEXT:    v_add_u32_e32 v3, 0x300, v3
; GFX9-NEXT:    v_add_u32_sdwa v4, v4, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s74, s74, 3
; GFX9-NEXT:    v_mov_b32_e32 v4, s73
; GFX9-NEXT:    s_add_i32 s72, s72, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s63
; GFX9-NEXT:    v_perm_b32 v4, s74, v4, v11
; GFX9-NEXT:    v_perm_b32 v5, s72, v5, v11
; GFX9-NEXT:    v_add_u32_e32 v4, 0x300, v4
; GFX9-NEXT:    v_add_u32_sdwa v5, v5, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v4, v4, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s62, s62, 3
; GFX9-NEXT:    v_mov_b32_e32 v5, s61
; GFX9-NEXT:    s_add_i32 s60, s60, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s59
; GFX9-NEXT:    v_perm_b32 v5, s62, v5, v11
; GFX9-NEXT:    v_perm_b32 v6, s60, v6, v11
; GFX9-NEXT:    v_add_u32_e32 v5, 0x300, v5
; GFX9-NEXT:    v_add_u32_sdwa v6, v6, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v5, v5, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s58, s58, 3
; GFX9-NEXT:    v_mov_b32_e32 v6, s57
; GFX9-NEXT:    s_add_i32 s56, s56, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s47
; GFX9-NEXT:    v_perm_b32 v6, s58, v6, v11
; GFX9-NEXT:    v_perm_b32 v7, s56, v7, v11
; GFX9-NEXT:    v_add_u32_e32 v6, 0x300, v6
; GFX9-NEXT:    v_add_u32_sdwa v7, v7, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    v_mov_b32_e32 v7, s45
; GFX9-NEXT:    s_add_i32 s44, s44, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s43
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v11
; GFX9-NEXT:    v_perm_b32 v8, s44, v8, v11
; GFX9-NEXT:    v_add_u32_e32 v7, 0x300, v7
; GFX9-NEXT:    v_add_u32_sdwa v8, v8, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v7, v7, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s42, s42, 3
; GFX9-NEXT:    v_mov_b32_e32 v8, s41
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s15
; GFX9-NEXT:    v_perm_b32 v8, s42, v8, v11
; GFX9-NEXT:    v_perm_b32 v9, s40, v9, v11
; GFX9-NEXT:    v_add_u32_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u32_sdwa v9, v9, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v8, v8, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    v_mov_b32_e32 v9, s13
; GFX9-NEXT:    s_add_i32 s12, s12, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s11
; GFX9-NEXT:    v_perm_b32 v9, s14, v9, v11
; GFX9-NEXT:    v_perm_b32 v10, s12, v10, v11
; GFX9-NEXT:    v_add_u32_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u32_sdwa v10, v10, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v9, v9, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    v_mov_b32_e32 v10, s9
; GFX9-NEXT:    s_add_i32 s8, s8, 3
; GFX9-NEXT:    v_mov_b32_e32 v12, s7
; GFX9-NEXT:    v_perm_b32 v10, s10, v10, v11
; GFX9-NEXT:    v_perm_b32 v12, s8, v12, v11
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v10, 0x300, v10
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v10, v10, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s5, 0xc0c0004
; GFX9-NEXT:    s_waitcnt vmcnt(18)
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v32
; GFX9-NEXT:    s_add_i32 s6, s6, 3
; GFX9-NEXT:    s_waitcnt vmcnt(10)
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v48
; GFX9-NEXT:    v_perm_b32 v14, v14, v39, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v14, v14, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v36
; GFX9-NEXT:    v_perm_b32 v15, v15, v35, s5
; GFX9-NEXT:    v_add_u32_sdwa v15, v15, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v60
; GFX9-NEXT:    s_waitcnt vmcnt(16)
; GFX9-NEXT:    v_perm_b32 v17, v17, v18, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v57
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v18, v18, v19, s5
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_perm_b32 v11, s6, v12, v11
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u32_sdwa v16, v16, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v17, v17, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_sdwa v18, v18, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v19
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v12
; GFX9-NEXT:    v_perm_b32 v12, v12, v13, s5
; GFX9-NEXT:    v_add_u32_sdwa v12, v12, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v11, v11, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v12, 3, v54
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v52
; GFX9-NEXT:    v_perm_b32 v12, v12, v53, s5
; GFX9-NEXT:    v_perm_b32 v13, v13, v51, s5
; GFX9-NEXT:    v_add_u32_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u32_sdwa v13, v13, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v12, v12, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v13, 3, v50
; GFX9-NEXT:    v_perm_b32 v13, v13, v49, s5
; GFX9-NEXT:    v_add_u32_e32 v13, 0x300, v13
; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v14, 3, v38
; GFX9-NEXT:    v_perm_b32 v14, v14, v37, s5
; GFX9-NEXT:    v_add_u32_e32 v14, 0x300, v14
; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v15, 3, v34
; GFX9-NEXT:    v_perm_b32 v15, v15, v33, s5
; GFX9-NEXT:    v_add_u32_e32 v15, 0x300, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v16, 3, v62
; GFX9-NEXT:    v_perm_b32 v16, v16, v61, s5
; GFX9-NEXT:    v_add_u32_e32 v16, 0x300, v16
; GFX9-NEXT:    v_or_b32_sdwa v16, v16, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v17, 3, v41
; GFX9-NEXT:    v_perm_b32 v17, v17, v40, s5
; GFX9-NEXT:    v_add_u32_e32 v17, 0x300, v17
; GFX9-NEXT:    v_or_b32_sdwa v17, v17, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v18, 3, v42
; GFX9-NEXT:    v_perm_b32 v18, v18, v47, s5
; GFX9-NEXT:    v_add_u32_e32 v18, 0x300, v18
; GFX9-NEXT:    v_add_u32_sdwa v19, v19, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v18, v18, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v19, 3, v46
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s5
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v19, 0x300, v19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v20
; GFX9-NEXT:    v_perm_b32 v20, v20, v21, s5
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v20, v20, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v19, v19, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v20, 3, v45
; GFX9-NEXT:    v_perm_b32 v20, v20, v44, s5
; GFX9-NEXT:    v_add_u32_e32 v20, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v21
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_sdwa v21, v21, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v20, v20, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    v_add_u32_e32 v21, 3, v43
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v21, v22, s5
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v21, 0x300, v21
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    v_add_u32_sdwa v22, v22, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v21, v21, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v22, 3, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v22, v23, s5
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v22, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    v_add_u32_sdwa v23, v23, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v22, v22, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v23, 3, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v23, v24, s5
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v23, 0x300, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    v_add_u32_sdwa v24, v24, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v23, v23, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v24, 3, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v24, v25, s5
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v24, 0x300, v24
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    v_add_u32_sdwa v25, v25, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v25, 3, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s5
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v25, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    v_add_u32_sdwa v26, v26, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v26, 3, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v26, v27, s5
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v26, 0x300, v26
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    v_add_u32_sdwa v27, v27, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s5
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v27, 0x300, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    v_add_u32_sdwa v28, v28, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v28, 3, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v28, v29, s5
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v28, 0x300, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    v_add_u32_sdwa v29, v29, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v29, 3, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s5
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v29, 0x300, v29
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    v_add_u32_sdwa v30, v30, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v30, 3, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v30, v31, s5
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v30, 0x300, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    v_add_u32_sdwa v31, v31, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v31, 3, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s5
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v31
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v32, 3, v32
; GFX9-NEXT:    v_perm_b32 v32, v32, v33, s5
; GFX9-NEXT:    v_add_u32_sdwa v32, v32, s4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
; GFX9-NEXT:  .LBB75_5: ; %end
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v16f64_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:260
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:188
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:60
; GFX11-TRUE16-NEXT:    s_clause 0xf
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s75, 0
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB75_4
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v135, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v55, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v147, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v99, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v103, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v67, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-TRUE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB75_3
; GFX11-TRUE16-NEXT:  .LBB75_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-TRUE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 3, v135
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 3, v99
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 3, v55
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-TRUE16-NEXT:  .LBB75_3: ; %end
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT:  .LBB75_4:
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_branch .LBB75_2
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v16f64_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v32, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v34, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v33, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v37, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v35, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v39, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v36, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v48, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v38, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v49, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v50, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v52, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v51, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:260
; GFX11-FAKE16-NEXT:    scratch_load_u16 v53, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v54, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v55, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:188
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:132
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:60
; GFX11-FAKE16-NEXT:    s_clause 0xf
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v30
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v29
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v28
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v27
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v26
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v25
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v24
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v23
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v22
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v21
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v20
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v19
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v18
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v17
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v16
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s74, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s75, 0
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-FAKE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB75_4
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v182, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v165, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v150, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v144, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v119, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v102, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v85, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v70, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s2, s3, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, s6, s5, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v64, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v8, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s42, s41, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v9, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v9, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v12, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v180, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v181, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v13, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v163, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v167, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v164, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v146, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v161, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v147, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v17, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v129, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v145, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v132, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v114, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v128, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v117, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v21, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v100, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v103, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v101, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v23, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v83, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v96, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v84, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v25, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v66, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v81, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v67, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v27, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v65, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v52, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v183
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v183, 16, v40
; GFX11-FAKE16-NEXT:    v_perm_b32 v40, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v41, 16, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v31, v183
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v40, v41
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB75_3
; GFX11-FAKE16-NEXT:  .LBB75_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s2, s3, v10
; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v1
; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s0, s1, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s16, s17, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s18, s19, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s24, s25, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s26, s27, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s72, s63, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v2, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s20, s21, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s22, s23, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s62, s61, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s56, s47, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v2, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v4, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s28, s29, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s74, s73, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s46, s45, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s40, s15, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v6, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s60, s59, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s58, s57, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s14, s13, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    s_add_i32 s10, s10, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    s_add_i32 s8, s8, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, s8, s7, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v8, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s44, s43, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s42, s41, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 3, v167
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v165
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v164
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v163
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v161
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v150
; GFX11-FAKE16-NEXT:    s_add_i32 s4, s4, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, s12, s11, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, s10, s9, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s6, s5, v10
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, s4, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v11, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 3, v182
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 3, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 3, v180
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v15, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v20, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 3, v147
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 3, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 3, v145
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 3, v144
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 3, v132
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v129
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v128
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v119
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v117
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v114
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v116, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, v23, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, v25, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 3, v103
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 3, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 3, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 3, v100
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 3, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v84
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v81
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v70
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v25
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v26
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, v22, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, v24, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v26, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v28, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v30, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 3, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 3, v66
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 3, v65
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 3, v64
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 3, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 3, v49
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, 3, v48
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 3, v39
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 3, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 3, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v38, v49, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v39, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v37, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v32, v34, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v30
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v31
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v36
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v29, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v31, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v36, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v33, v32
; GFX11-FAKE16-NEXT:  .LBB75_3: ; %end
; GFX11-FAKE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-FAKE16-NEXT:  .LBB75_4:
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_branch .LBB75_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define <64 x bfloat> @bitcast_v16f64_to_v64bf16(<16 x double> %a, i32 %b) {
; SI-LABEL: bitcast_v16f64_to_v64bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB76_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v14
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v13
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v12
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v11
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v10
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v9
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v8
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v7
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v6
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v5
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v4
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v3
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v2
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v1
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:  .LBB76_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB76_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v14
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v13
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v12
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v11
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v10
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v9
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v8
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v7
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v6
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v5
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v4
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v3
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v2
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v1
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v0
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:  .LBB76_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v63
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v62
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v60
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v58
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v56
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v46
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v44
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v42
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v40
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v54
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v52
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v50
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v48
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v39
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v38
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v37
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v36
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v35
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v34
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v33
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_alignbit_b32 v18, v18, v19, 16
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_alignbit_b32 v19, v19, v20, 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_alignbit_b32 v21, v21, v22, 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_alignbit_b32 v22, v22, v23, 16
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_alignbit_b32 v23, v23, v24, 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_alignbit_b32 v24, v24, v25, 16
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_alignbit_b32 v25, v25, v26, 16
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_alignbit_b32 v26, v26, v27, 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_alignbit_b32 v27, v27, v28, 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_alignbit_b32 v28, v28, v29, 16
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_alignbit_b32 v29, v29, v30, 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_alignbit_b32 v30, v30, v31, 16
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; SI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v64bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB76_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; VI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; VI-NEXT:  .LBB76_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v64bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB76_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX9-NEXT:  .LBB76_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v64bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB76_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX11-NEXT:  .LBB76_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define inreg <64 x bfloat> @bitcast_v16f64_to_v64bf16_scalar(<16 x double> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16f64_to_v64bf16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_writelane_b32 v63, s70, 22
; SI-NEXT:    v_writelane_b32 v63, s71, 23
; SI-NEXT:    v_writelane_b32 v63, s80, 24
; SI-NEXT:    v_writelane_b32 v63, s81, 25
; SI-NEXT:    v_writelane_b32 v63, s82, 26
; SI-NEXT:    v_writelane_b32 v63, s83, 27
; SI-NEXT:    v_writelane_b32 v63, s84, 28
; SI-NEXT:    v_writelane_b32 v63, s85, 29
; SI-NEXT:    v_writelane_b32 v63, s86, 30
; SI-NEXT:    v_writelane_b32 v63, s87, 31
; SI-NEXT:    v_writelane_b32 v63, s96, 32
; SI-NEXT:    v_writelane_b32 v63, s97, 33
; SI-NEXT:    v_writelane_b32 v63, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s44, v18
; SI-NEXT:    v_writelane_b32 v63, s99, 35
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_readfirstlane_b32 s7, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    v_readfirstlane_b32 s9, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_readfirstlane_b32 s10, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s15, v7
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s40, v4
; SI-NEXT:    v_readfirstlane_b32 s43, v3
; SI-NEXT:    v_readfirstlane_b32 s42, v2
; SI-NEXT:    v_readfirstlane_b32 s45, v1
; SI-NEXT:    s_cmp_lg_u32 s44, 0
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB77_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s46, s17, 0xffff0000
; SI-NEXT:    v_writelane_b32 v62, s46, 0
; SI-NEXT:    s_lshl_b32 s46, s17, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 1
; SI-NEXT:    s_and_b32 s46, s16, 0xffff0000
; SI-NEXT:    v_writelane_b32 v62, s46, 2
; SI-NEXT:    s_lshl_b32 s46, s16, 16
; SI-NEXT:    s_and_b32 s59, s5, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s5, 16
; SI-NEXT:    s_and_b32 s57, s4, 0xffff0000
; SI-NEXT:    s_lshl_b32 s56, s4, 16
; SI-NEXT:    s_and_b32 s99, s7, 0xffff0000
; SI-NEXT:    s_lshl_b32 s98, s7, 16
; SI-NEXT:    s_and_b32 s97, s6, 0xffff0000
; SI-NEXT:    s_lshl_b32 s96, s6, 16
; SI-NEXT:    s_and_b32 s87, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s86, s9, 16
; SI-NEXT:    s_and_b32 s85, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s84, s8, 16
; SI-NEXT:    s_and_b32 s83, s11, 0xffff0000
; SI-NEXT:    s_lshl_b32 s82, s11, 16
; SI-NEXT:    s_and_b32 s81, s10, 0xffff0000
; SI-NEXT:    s_lshl_b32 s80, s10, 16
; SI-NEXT:    s_and_b32 s71, s13, 0xffff0000
; SI-NEXT:    s_lshl_b32 s70, s13, 16
; SI-NEXT:    s_and_b32 s69, s12, 0xffff0000
; SI-NEXT:    s_lshl_b32 s68, s12, 16
; SI-NEXT:    s_and_b32 s67, s15, 0xffff0000
; SI-NEXT:    s_lshl_b32 s66, s15, 16
; SI-NEXT:    s_and_b32 s65, s14, 0xffff0000
; SI-NEXT:    s_lshl_b32 s64, s14, 16
; SI-NEXT:    s_and_b32 s55, s41, 0xffff0000
; SI-NEXT:    s_lshl_b32 s54, s41, 16
; SI-NEXT:    s_and_b32 s53, s40, 0xffff0000
; SI-NEXT:    s_lshl_b32 s52, s40, 16
; SI-NEXT:    s_and_b32 s51, s43, 0xffff0000
; SI-NEXT:    s_lshl_b32 s50, s43, 16
; SI-NEXT:    s_and_b32 s49, s42, 0xffff0000
; SI-NEXT:    s_lshl_b32 s48, s42, 16
; SI-NEXT:    s_and_b32 s39, s45, 0xffff0000
; SI-NEXT:    s_lshl_b32 s38, s45, 16
; SI-NEXT:    s_and_b32 s37, s44, 0xffff0000
; SI-NEXT:    s_lshl_b32 s36, s44, 16
; SI-NEXT:    s_and_b32 s35, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s34, s29, 16
; SI-NEXT:    s_and_b32 s31, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s30, s28, 16
; SI-NEXT:    s_and_b32 s95, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s27, 16
; SI-NEXT:    s_and_b32 s93, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s26, 16
; SI-NEXT:    s_and_b32 s91, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s25, 16
; SI-NEXT:    s_and_b32 s89, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s24, 16
; SI-NEXT:    s_and_b32 s79, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s78, s23, 16
; SI-NEXT:    s_and_b32 s77, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s76, s22, 16
; SI-NEXT:    s_and_b32 s75, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s21, 16
; SI-NEXT:    s_and_b32 s73, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s20, 16
; SI-NEXT:    s_and_b32 s63, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s19, 16
; SI-NEXT:    s_and_b32 s61, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s18, 16
; SI-NEXT:    v_writelane_b32 v62, s46, 3
; SI-NEXT:    s_cbranch_execnz .LBB77_4
; SI-NEXT:  .LBB77_2: ; %cmp.true
; SI-NEXT:    v_add_f64 v[22:23], s[4:5], 1.0
; SI-NEXT:    v_add_f64 v[0:1], s[22:23], 1.0
; SI-NEXT:    v_add_f64 v[2:3], s[24:25], 1.0
; SI-NEXT:    v_add_f64 v[14:15], s[14:15], 1.0
; SI-NEXT:    v_add_f64 v[16:17], s[12:13], 1.0
; SI-NEXT:    v_add_f64 v[24:25], s[6:7], 1.0
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_add_f64 v[10:11], s[42:43], 1.0
; SI-NEXT:    v_add_f64 v[18:19], s[10:11], 1.0
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v17
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v15
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v14
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v2
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v1
; SI-NEXT:    v_add_f64 v[1:2], s[20:21], 1.0
; SI-NEXT:    v_add_f64 v[20:21], s[8:9], 1.0
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v25
; SI-NEXT:    v_add_f64 v[4:5], s[26:27], 1.0
; SI-NEXT:    v_add_f64 v[6:7], s[28:29], 1.0
; SI-NEXT:    v_add_f64 v[8:9], s[44:45], 1.0
; SI-NEXT:    v_add_f64 v[12:13], s[40:41], 1.0
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v24
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v19
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v18
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v11
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v3
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v2
; SI-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; SI-NEXT:    v_add_f64 v[37:38], s[16:17], 1.0
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v25
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v24
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v20
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v16
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v13
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v12
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v9
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v8
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v7
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v7
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v6
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v5
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v4
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v0
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v3
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v38
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v38
; SI-NEXT:    v_mov_b32_e32 v38, v34
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v37
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v37
; SI-NEXT:    s_branch .LBB77_5
; SI-NEXT:  .LBB77_3:
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr61
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr63
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr73
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr75
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr77
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr93
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr31
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr35
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr37
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr71
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr81
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr83
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr85
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr87
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr97
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr99
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr57
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr59
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; kill: killed $sgpr46
; SI-NEXT:    s_branch .LBB77_2
; SI-NEXT:  .LBB77_4:
; SI-NEXT:    v_mov_b32_e32 v0, s59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s56
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s99
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s97
; SI-NEXT:    v_readlane_b32 s4, v62, 0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s96
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s86
; SI-NEXT:    v_mov_b32_e32 v9, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 3
; SI-NEXT:    v_mov_b32_e32 v38, s98
; SI-NEXT:    v_mov_b32_e32 v51, s87
; SI-NEXT:    v_mov_b32_e32 v53, s85
; SI-NEXT:    v_mov_b32_e32 v52, s84
; SI-NEXT:    v_mov_b32_e32 v55, s83
; SI-NEXT:    v_mov_b32_e32 v54, s82
; SI-NEXT:    v_mov_b32_e32 v41, s81
; SI-NEXT:    v_mov_b32_e32 v40, s80
; SI-NEXT:    v_mov_b32_e32 v43, s71
; SI-NEXT:    v_mov_b32_e32 v42, s70
; SI-NEXT:    v_mov_b32_e32 v45, s69
; SI-NEXT:    v_mov_b32_e32 v44, s68
; SI-NEXT:    v_mov_b32_e32 v47, s67
; SI-NEXT:    v_mov_b32_e32 v46, s66
; SI-NEXT:    v_mov_b32_e32 v57, s65
; SI-NEXT:    v_mov_b32_e32 v56, s64
; SI-NEXT:    v_mov_b32_e32 v59, s55
; SI-NEXT:    v_mov_b32_e32 v58, s54
; SI-NEXT:    v_mov_b32_e32 v61, s53
; SI-NEXT:    v_mov_b32_e32 v60, s52
; SI-NEXT:    v_mov_b32_e32 v48, s51
; SI-NEXT:    v_mov_b32_e32 v39, s50
; SI-NEXT:    v_mov_b32_e32 v33, s49
; SI-NEXT:    v_mov_b32_e32 v32, s48
; SI-NEXT:    v_mov_b32_e32 v31, s39
; SI-NEXT:    v_mov_b32_e32 v30, s38
; SI-NEXT:    v_mov_b32_e32 v29, s37
; SI-NEXT:    v_mov_b32_e32 v28, s36
; SI-NEXT:    v_mov_b32_e32 v27, s35
; SI-NEXT:    v_mov_b32_e32 v26, s34
; SI-NEXT:    v_mov_b32_e32 v25, s31
; SI-NEXT:    v_mov_b32_e32 v24, s30
; SI-NEXT:    v_mov_b32_e32 v23, s95
; SI-NEXT:    v_mov_b32_e32 v22, s94
; SI-NEXT:    v_mov_b32_e32 v21, s93
; SI-NEXT:    v_mov_b32_e32 v20, s92
; SI-NEXT:    v_mov_b32_e32 v19, s91
; SI-NEXT:    v_mov_b32_e32 v18, s90
; SI-NEXT:    v_mov_b32_e32 v17, s89
; SI-NEXT:    v_mov_b32_e32 v50, s88
; SI-NEXT:    v_mov_b32_e32 v15, s79
; SI-NEXT:    v_mov_b32_e32 v14, s78
; SI-NEXT:    v_mov_b32_e32 v12, s77
; SI-NEXT:    v_mov_b32_e32 v36, s76
; SI-NEXT:    v_mov_b32_e32 v11, s75
; SI-NEXT:    v_mov_b32_e32 v10, s74
; SI-NEXT:    v_mov_b32_e32 v49, s73
; SI-NEXT:    v_mov_b32_e32 v13, s72
; SI-NEXT:    v_mov_b32_e32 v1, s63
; SI-NEXT:    v_mov_b32_e32 v6, s62
; SI-NEXT:    v_mov_b32_e32 v5, s61
; SI-NEXT:    v_mov_b32_e32 v4, s60
; SI-NEXT:    v_mov_b32_e32 v3, s4
; SI-NEXT:  .LBB77_5: ; %end
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v2
; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v13
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v9
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v4
; SI-NEXT:    v_lshr_b64 v[33:34], v[34:35], 16
; SI-NEXT:    v_lshr_b64 v[34:35], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[2:3], v[7:8], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v39
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v60
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v58
; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v46
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v44
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v42
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v40
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v38
; SI-NEXT:    v_readlane_b32 s99, v63, 35
; SI-NEXT:    v_readlane_b32 s98, v63, 34
; SI-NEXT:    v_readlane_b32 s97, v63, 33
; SI-NEXT:    v_readlane_b32 s96, v63, 32
; SI-NEXT:    v_readlane_b32 s87, v63, 31
; SI-NEXT:    v_readlane_b32 s86, v63, 30
; SI-NEXT:    v_readlane_b32 s85, v63, 29
; SI-NEXT:    v_readlane_b32 s84, v63, 28
; SI-NEXT:    v_readlane_b32 s83, v63, 27
; SI-NEXT:    v_readlane_b32 s82, v63, 26
; SI-NEXT:    v_readlane_b32 s81, v63, 25
; SI-NEXT:    v_readlane_b32 s80, v63, 24
; SI-NEXT:    v_readlane_b32 s71, v63, 23
; SI-NEXT:    v_readlane_b32 s70, v63, 22
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    v_lshr_b64 v[4:5], v[5:6], 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[7:8], v[15:16], 16
; SI-NEXT:    v_lshr_b64 v[8:9], v[9:10], 16
; SI-NEXT:    v_lshr_b64 v[9:10], v[11:12], 16
; SI-NEXT:    v_lshr_b64 v[10:11], v[13:14], 16
; SI-NEXT:    v_lshr_b64 v[11:12], v[22:23], 16
; SI-NEXT:    v_lshr_b64 v[12:13], v[24:25], 16
; SI-NEXT:    v_lshr_b64 v[13:14], v[26:27], 16
; SI-NEXT:    v_lshr_b64 v[14:15], v[28:29], 16
; SI-NEXT:    v_lshr_b64 v[15:16], v[30:31], 16
; SI-NEXT:    v_lshr_b64 v[16:17], v[17:18], 16
; SI-NEXT:    v_lshr_b64 v[17:18], v[19:20], 16
; SI-NEXT:    v_lshr_b64 v[18:19], v[36:37], 16
; SI-NEXT:    v_lshr_b64 v[19:20], v[58:59], 16
; SI-NEXT:    v_lshr_b64 v[20:21], v[56:57], 16
; SI-NEXT:    v_lshr_b64 v[21:22], v[46:47], 16
; SI-NEXT:    v_lshr_b64 v[22:23], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[23:24], v[42:43], 16
; SI-NEXT:    v_lshr_b64 v[24:25], v[40:41], 16
; SI-NEXT:    v_lshr_b64 v[25:26], v[54:55], 16
; SI-NEXT:    v_lshr_b64 v[26:27], v[52:53], 16
; SI-NEXT:    v_lshr_b64 v[27:28], v[50:51], 16
; SI-NEXT:    v_lshr_b64 v[28:29], v[48:49], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[29:30], v[38:39], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[30:31], v[60:61], 16
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshr_b64 v[31:32], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v0, v33
; SI-NEXT:    v_mov_b32_e32 v1, v34
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v64bf16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s36, 2
; VI-NEXT:    v_writelane_b32 v32, s37, 3
; VI-NEXT:    v_writelane_b32 v32, s38, 4
; VI-NEXT:    v_writelane_b32 v32, s39, 5
; VI-NEXT:    v_writelane_b32 v32, s48, 6
; VI-NEXT:    v_writelane_b32 v32, s49, 7
; VI-NEXT:    v_writelane_b32 v32, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB77_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB77_4
; VI-NEXT:  .LBB77_2: ; %cmp.true
; VI-NEXT:    v_add_f64 v[14:15], s[30:31], 1.0
; VI-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; VI-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; VI-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; VI-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; VI-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; VI-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; VI-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; VI-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; VI-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; VI-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; VI-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; VI-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; VI-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; VI-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; VI-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; VI-NEXT:    s_branch .LBB77_5
; VI-NEXT:  .LBB77_3:
; VI-NEXT:    s_branch .LBB77_2
; VI-NEXT:  .LBB77_4:
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:  .LBB77_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v32, 9
; VI-NEXT:    v_readlane_b32 s50, v32, 8
; VI-NEXT:    v_readlane_b32 s49, v32, 7
; VI-NEXT:    v_readlane_b32 s48, v32, 6
; VI-NEXT:    v_readlane_b32 s39, v32, 5
; VI-NEXT:    v_readlane_b32 s38, v32, 4
; VI-NEXT:    v_readlane_b32 s37, v32, 3
; VI-NEXT:    v_readlane_b32 s36, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v64bf16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB77_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB77_4
; GFX9-NEXT:  .LBB77_2: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[14:15], s[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; GFX9-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; GFX9-NEXT:    s_branch .LBB77_5
; GFX9-NEXT:  .LBB77_3:
; GFX9-NEXT:    s_branch .LBB77_2
; GFX9-NEXT:  .LBB77_4:
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:  .LBB77_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v64bf16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB77_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB77_4
; GFX11-NEXT:  .LBB77_2: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[14:15], s[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], s[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], s[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], s[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], s[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], s[16:17], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], s[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], s[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; GFX11-NEXT:    s_branch .LBB77_5
; GFX11-NEXT:  .LBB77_3:
; GFX11-NEXT:    s_branch .LBB77_2
; GFX11-NEXT:  .LBB77_4:
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:  .LBB77_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define <16 x double> @bitcast_v64bf16_to_v16f64(<64 x bfloat> %a, i32 %b) {
; SI-LABEL: bitcast_v64bf16_to_v16f64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v0
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v63
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v62
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v60
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v58, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_and_b32_e32 v46, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v36, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v53
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v37
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v38
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v38
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v42
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v40
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v55
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v18
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v19
; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v51
; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v20
; SI-NEXT:    v_mul_f32_e32 v45, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v49
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v41, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v23
; SI-NEXT:    v_mul_f32_e32 v55, 1.0, v39
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v53, 1.0, v36
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v25
; SI-NEXT:    v_mul_f32_e32 v51, 1.0, v35
; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v49, 1.0, v34
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v39, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v29
; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v31
; SI-NEXT:    v_mul_f32_e32 v35, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v33, 1.0, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB78_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v51
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v39
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v36
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v34
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    v_alignbit_b32 v19, v19, v59, 16
; SI-NEXT:    v_alignbit_b32 v20, v20, v57, 16
; SI-NEXT:    v_alignbit_b32 v21, v21, v46, 16
; SI-NEXT:    v_alignbit_b32 v22, v22, v44, 16
; SI-NEXT:    v_alignbit_b32 v23, v23, v42, 16
; SI-NEXT:    v_alignbit_b32 v24, v24, v40, 16
; SI-NEXT:    v_alignbit_b32 v25, v25, v54, 16
; SI-NEXT:    v_alignbit_b32 v26, v26, v52, 16
; SI-NEXT:    v_alignbit_b32 v27, v27, v50, 16
; SI-NEXT:    v_alignbit_b32 v28, v28, v48, 16
; SI-NEXT:    v_alignbit_b32 v29, v29, v38, 16
; SI-NEXT:    v_alignbit_b32 v30, v30, v35, 16
; SI-NEXT:    v_alignbit_b32 v31, v31, v33, 16
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_alignbit_b32 v0, v0, v63, 16
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v62
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v60
; SI-NEXT:    v_alignbit_b32 v18, v18, v61, 16
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:  .LBB78_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB78_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v63
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v60
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v55
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v0, v1, v0, 16
; SI-NEXT:    v_alignbit_b32 v1, v3, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v2, v3, v2, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_alignbit_b32 v3, v4, v3, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_alignbit_b32 v4, v5, v4, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_alignbit_b32 v5, v6, v5, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_alignbit_b32 v6, v7, v6, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_alignbit_b32 v7, v8, v7, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_alignbit_b32 v8, v9, v8, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_alignbit_b32 v9, v10, v9, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_alignbit_b32 v10, v11, v10, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_alignbit_b32 v11, v12, v11, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_alignbit_b32 v12, v13, v12, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v13, v14, v13, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_alignbit_b32 v14, v15, v14, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v15, v16, v15, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v16, v17, v16, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_alignbit_b32 v17, v18, v17, 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v61
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_alignbit_b32 v18, v19, v18, 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v59
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_alignbit_b32 v19, v20, v19, 16
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v57
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_alignbit_b32 v20, v21, v20, 16
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_alignbit_b32 v21, v22, v21, 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v44
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_alignbit_b32 v22, v23, v22, 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_alignbit_b32 v23, v24, v23, 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_alignbit_b32 v24, v25, v24, 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_alignbit_b32 v25, v26, v25, 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_alignbit_b32 v26, v27, v26, 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_alignbit_b32 v27, v28, v27, 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_alignbit_b32 v28, v29, v28, 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_alignbit_b32 v29, v30, v29, 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v35
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_alignbit_b32 v30, v31, v30, 16
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v33
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_alignbit_b32 v31, v32, v31, 16
; SI-NEXT:  .LBB78_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v16f64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB78_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v15, 16, 1
; VI-NEXT:    s_movk_i32 s6, 0x7fff
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v15
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_alignbit_b32 v15, v15, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v14, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v14
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; VI-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_alignbit_b32 v14, v14, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v13
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_alignbit_b32 v13, v13, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v12, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v12
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; VI-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_alignbit_b32 v12, v12, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v11
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_alignbit_b32 v11, v11, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v10, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v10
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; VI-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_alignbit_b32 v10, v10, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v9
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_alignbit_b32 v9, v9, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v8, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v8
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; VI-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_alignbit_b32 v8, v8, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v7
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_alignbit_b32 v7, v7, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v6, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v6
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; VI-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_alignbit_b32 v6, v6, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v5
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_alignbit_b32 v5, v5, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v4, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v4
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; VI-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_alignbit_b32 v4, v4, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v3
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_alignbit_b32 v3, v3, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v2
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_alignbit_b32 v2, v2, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v1
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_alignbit_b32 v1, v1, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v0
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; VI-NEXT:    v_alignbit_b32 v0, v0, v32, 16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v30, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v30
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; VI-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_alignbit_b32 v30, v30, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v29, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v29
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_alignbit_b32 v29, v29, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v28, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v28
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; VI-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_alignbit_b32 v28, v28, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v27
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_alignbit_b32 v27, v27, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v26, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v26
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; VI-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_alignbit_b32 v26, v26, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v25
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_alignbit_b32 v25, v25, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v24, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v24
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; VI-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_alignbit_b32 v24, v24, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v23
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_alignbit_b32 v23, v23, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v22, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v22
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; VI-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_alignbit_b32 v22, v22, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v21
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_alignbit_b32 v21, v21, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v20, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v20
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; VI-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_alignbit_b32 v20, v20, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v19
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_alignbit_b32 v19, v19, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v18
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_alignbit_b32 v18, v18, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v17
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_alignbit_b32 v17, v17, v32, 16
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v16
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_alignbit_b32 v16, v16, v32, 16
; VI-NEXT:  .LBB78_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v16f64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB78_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    s_movk_i32 s6, 0x7fff
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v15, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v15, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v15
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v33, v34, vcc
; GFX9-NEXT:    s_mov_b32 s7, 0x7060302
; GFX9-NEXT:    v_perm_b32 v15, v15, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v14
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v14, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v14, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v14, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v14, v14, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v13
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v13, v13, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v12
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v12, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v12, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v12
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v12, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v12, v12, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v11
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v11, v11, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v10, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v10, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v10
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v10, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v10, v10, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v9
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v9, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v9
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v9, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v9, v9, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v8
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v8, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v8, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v8
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v8, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v8, v8, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v7
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v7, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v7
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v7, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v7, v7, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v6, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v6, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v6, v6, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v5, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v5, v5, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v4
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v4, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v4, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v4, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v4, v4, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v3, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v3, v3, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v2, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v2, v2, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v1, v1, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v0
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v0, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v0, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v0, v0, v32, s7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v31, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v30
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v30, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v30, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v30
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v30, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v30, v30, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v29
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v29, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v29
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v29, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v29, v29, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v28
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v28, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v28, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v28
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v28, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v28, v28, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v27
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v27, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v27
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v27, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v27, v27, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v26
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v26, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v26, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v26
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v26, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v26, v26, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v25
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v25, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v25
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v25, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v25, v25, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v24
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v24, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v24, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v24
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v24, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v24, v24, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v23
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v23, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v23
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v23, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v23, v23, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v22
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v22, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v22
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v22, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v22, v22, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v21
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v21, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v21, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v21, v21, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v20, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v20, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v20, v20, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v19
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v19, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v19, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v19, v19, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v18, v18, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v17, v17, v32, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v16, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; GFX9-NEXT:    v_perm_b32 v16, v16, v32, s7
; GFX9-NEXT:  .LBB78_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v16f64:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47_vgpr48_vgpr49_vgpr50_vgpr51_vgpr52_vgpr53_vgpr54_vgpr55_vgpr56_vgpr57_vgpr58_vgpr59_vgpr60_vgpr61_vgpr62_vgpr63
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB78_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v63, v31 :: v_dual_mov_b32 v62, v30
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v61, v29 :: v_dual_mov_b32 v60, v28
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v59, v27 :: v_dual_mov_b32 v58, v26
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v57, v25 :: v_dual_mov_b32 v56, v24
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v55, v23 :: v_dual_mov_b32 v54, v22
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v53, v21 :: v_dual_mov_b32 v52, v20
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, v19 :: v_dual_mov_b32 v50, v18
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v49, v17 :: v_dual_mov_b32 v48, v16
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v47, v15 :: v_dual_mov_b32 v46, v14
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v45, v13 :: v_dual_mov_b32 v44, v12
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v43, v11 :: v_dual_mov_b32 v42, v10
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v41, v9 :: v_dual_mov_b32 v40, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v7 :: v_dual_mov_b32 v38, v6
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v5 :: v_dual_mov_b32 v36, v4
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v3 :: v_dual_mov_b32 v34, v2
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v1 :: v_dual_mov_b32 v32, v0
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT:  .LBB78_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB78_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v32, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v32, v35, v38 :: v_dual_and_b32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v39, v33, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v47, v36, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v32.h
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_add3_u32 v38, v48, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v38, v49, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v15, 0x40c00000, v34 :: v_dual_cndmask_b32 v46, v35, v32
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, v14.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v33, v34, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v13, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v45, v33, v35
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v13.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v14, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v15, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v44, v32, v34, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, v12.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v15, v15, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v11, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v14, v32 :: v_dual_and_b32 v12, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v43, v15, v33
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, v11.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v14, v14, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v10, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v13, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v42, v14, v32, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v10.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v13, v13, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v9, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v41, v13, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, v9.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v12, v12, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v8, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v11, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v40, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v8.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v11, v11, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v7, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v39, v11, v13
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v7.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v10, v10, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v6, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v9, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v6.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v9, v9, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v5, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v8, v10 :: v_dual_and_b32 v6, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v37, v9, v11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v5.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v8, v8, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v7, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v8, v10, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v4.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v3, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v6, v8 :: v_dual_and_b32 v4, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_cndmask_b32 v35, v7, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v5, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v6, v8, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v5, v5, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v4, v6 :: v_dual_and_b32 v2, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_cndmask_b32 v33, v5, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v1.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v31
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v5, 16, v31
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v3, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v4, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v0.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v3, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v63, v3, v7 :: v_dual_and_b32 v0, 0xffff0000, v30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v29
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v62, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v28
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v61, v3, v7 :: v_dual_lshlrev_b32 v4, 16, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v61.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v60, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v59, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v59.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v25
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_lshlrev_b32 v4, 16, v26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v58, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v24
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v57, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, v2.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v56, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v56.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v4, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v55, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.l, v2.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v4, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_lshlrev_b32 v5, 16, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v53, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v52, v0, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v4, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.l, v2.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_lshlrev_b32 v6, 16, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v4, 0x40c00000, v6 :: v_dual_cndmask_b32 v51, v0, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v3.h
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v8, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v50, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v5, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v16
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v1 :: v_dual_add_f32 v1, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v0.h
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v0, v0, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v9, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v6, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.l, v0.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v1, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.l, v5.h
; GFX11-TRUE16-NEXT:  .LBB78_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v40 :: v_dual_mov_b32 v9, v41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v42 :: v_dual_mov_b32 v11, v43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, v44 :: v_dual_mov_b32 v13, v45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v46 :: v_dual_mov_b32 v15, v47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v56 :: v_dual_mov_b32 v25, v57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v58 :: v_dual_mov_b32 v27, v59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v60 :: v_dual_mov_b32 v29, v61
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v62 :: v_dual_mov_b32 v31, v63
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v33
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v34 :: v_dual_mov_b32 v3, v35
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v36 :: v_dual_mov_b32 v5, v37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v38 :: v_dual_mov_b32 v7, v39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v48 :: v_dual_mov_b32 v17, v49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v50 :: v_dual_mov_b32 v19, v51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v52 :: v_dual_mov_b32 v21, v53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v54 :: v_dual_mov_b32 v23, v55
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v16f64:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB78_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v32
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v39, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v15, v36, v38 :: v_dual_lshlrev_b32 v36, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v33, v37, v34 :: v_dual_add_f32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v13, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_cndmask_b32 v32, v32, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v13, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v11, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v35, v38 :: v_dual_lshlrev_b32 v37, 16, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_lshlrev_b32 v35, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v8, 0x40c00000, v8
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v9, 0x40c00000, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v8, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v9, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v32, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v34 :: v_dual_add_f32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v6, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v7, 0x40c00000, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v6, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v7, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v33, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v32, v36 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v5, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v37 :: v_dual_cndmask_b32 v4, v34, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v3, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v0, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v1, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v31
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v31, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v31, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v29
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v30, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v31, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v30, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v29, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v29, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v28, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v27
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v28, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v29, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v28, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v27, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v25
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v26, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v26, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v27, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v34, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v26, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v25, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v25, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v23
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v34, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v33, v23, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v37 :: v_dual_cndmask_b32 v33, v33, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v23, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v22, v35, v38 :: v_dual_lshlrev_b32 v35, 16, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v22, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v35
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v36, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v32, v21, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v32, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v38, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v36 :: v_dual_lshlrev_b32 v36, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v21, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v32, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v20, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v20, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v37, v19, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v48, v18, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v19, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v37, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v39
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v36, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v36
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v18, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v39, v17, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v16, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v17, v35, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v37, v50, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v16, v36, 0x7060302
; GFX11-FAKE16-NEXT:  .LBB78_2: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define inreg <16 x double> @bitcast_v64bf16_to_v16f64_scalar(<64 x bfloat> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64bf16_to_v16f64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    s_and_b32 s12, s25, 0xffff0000
; SI-NEXT:    s_and_b32 s30, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v1
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; SI-NEXT:    s_and_b32 s34, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_readfirstlane_b32 s94, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s34
; SI-NEXT:    v_readfirstlane_b32 s90, v5
; SI-NEXT:    s_and_b32 s95, s94, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s30
; SI-NEXT:    v_readfirstlane_b32 s88, v6
; SI-NEXT:    s_and_b32 s91, s90, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s95
; SI-NEXT:    v_readfirstlane_b32 s78, v7
; SI-NEXT:    s_and_b32 s89, s88, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s91
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    s_and_b32 s79, s78, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s89
; SI-NEXT:    v_readfirstlane_b32 s74, v9
; SI-NEXT:    s_and_b32 s77, s76, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s79
; SI-NEXT:    v_readfirstlane_b32 s72, v10
; SI-NEXT:    s_and_b32 s75, s74, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s77
; SI-NEXT:    v_readfirstlane_b32 s62, v11
; SI-NEXT:    s_and_b32 s73, s72, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s75
; SI-NEXT:    v_readfirstlane_b32 s60, v12
; SI-NEXT:    s_and_b32 s63, s62, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s73
; SI-NEXT:    v_readfirstlane_b32 s58, v13
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s63
; SI-NEXT:    v_readfirstlane_b32 s56, v14
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s61
; SI-NEXT:    v_readfirstlane_b32 s46, v15
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s59
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s57
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s47
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    v_readfirstlane_b32 s92, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s45
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s29, 16
; SI-NEXT:    s_lshl_b32 s7, s28, 16
; SI-NEXT:    s_and_b32 s8, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s9, s27, 16
; SI-NEXT:    s_and_b32 s10, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s11, s26, 16
; SI-NEXT:    s_lshl_b32 s13, s25, 16
; SI-NEXT:    s_and_b32 s14, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s15, s24, 16
; SI-NEXT:    s_and_b32 s24, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s25, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s26, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s27, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s28, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s29, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_and_b32 s41, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_and_b32 s93, s92, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_and_b32 s36, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s43
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cmp_lg_u32 vcc_lo, 0
; SI-NEXT:    v_mul_f32_e64 v55, 1.0, s41
; SI-NEXT:    v_mul_f32_e64 v53, 1.0, s40
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s29
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v62, 1.0, s27
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s26
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s25
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s24
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s14
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s10
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s8
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s4
; SI-NEXT:    v_mul_f32_e64 v39, 1.0, s36
; SI-NEXT:    v_mul_f32_e64 v47, 1.0, s93
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e64 v43, 1.0, s16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s17
; SI-NEXT:    v_mul_f32_e64 v34, 1.0, s18
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s19
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s20
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s21
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s22
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s15
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s13
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s11
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s9
; SI-NEXT:    v_mul_f32_e64 v61, 1.0, s7
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s5
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s37
; SI-NEXT:    v_mul_f32_e64 v60, 1.0, s35
; SI-NEXT:    v_mul_f32_e64 v33, 1.0, s31
; SI-NEXT:    v_mul_f32_e64 v57, 1.0, s94
; SI-NEXT:    v_mul_f32_e64 v59, 1.0, s92
; SI-NEXT:    v_mul_f32_e64 v56, 1.0, s90
; SI-NEXT:    v_mul_f32_e64 v38, 1.0, s88
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s78
; SI-NEXT:    v_mul_f32_e64 v36, 1.0, s76
; SI-NEXT:    v_mul_f32_e64 v46, 1.0, s74
; SI-NEXT:    v_mul_f32_e64 v48, 1.0, s72
; SI-NEXT:    v_mul_f32_e64 v49, 1.0, s62
; SI-NEXT:    v_mul_f32_e64 v45, 1.0, s60
; SI-NEXT:    v_mul_f32_e64 v42, 1.0, s58
; SI-NEXT:    v_mul_f32_e64 v51, 1.0, s56
; SI-NEXT:    v_mul_f32_e64 v41, 1.0, s46
; SI-NEXT:    v_mul_f32_e64 v54, 1.0, s44
; SI-NEXT:    v_mul_f32_e64 v52, 1.0, s42
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB79_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v55
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v53
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v44, v1
; SI-NEXT:    v_mov_b32_e32 v43, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v12
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v37, v36
; SI-NEXT:    v_mov_b32_e32 v36, v47
; SI-NEXT:    v_mov_b32_e32 v50, v49
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v40, v41
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v1, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[43:44], 16
; SI-NEXT:    v_mov_b32_e32 v2, v34
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[2:3], v[34:35], 16
; SI-NEXT:    v_mov_b32_e32 v3, v17
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[3:4], v[17:18], 16
; SI-NEXT:    v_mov_b32_e32 v4, v20
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[4:5], v[20:21], 16
; SI-NEXT:    v_mov_b32_e32 v5, v19
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v6
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[5:6], v[19:20], 16
; SI-NEXT:    v_mov_b32_e32 v6, v13
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[6:7], v[13:14], 16
; SI-NEXT:    v_mov_b32_e32 v7, v31
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[7:8], v[31:32], 16
; SI-NEXT:    v_mov_b32_e32 v8, v23
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[8:9], v[23:24], 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v11
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v44, v45
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v9
; SI-NEXT:    v_mov_b32_e32 v9, v25
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[9:10], v[25:26], 16
; SI-NEXT:    v_mov_b32_e32 v10, v30
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[10:11], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v11, v27
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[11:12], v[27:28], 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v18
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v12
; SI-NEXT:    v_mov_b32_e32 v12, v61
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[12:13], v[61:62], 16
; SI-NEXT:    v_mov_b32_e32 v13, v29
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[13:14], v[29:30], 16
; SI-NEXT:    v_mov_b32_e32 v14, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v62, v39
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v62
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[14:15], v[15:16], 16
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v15
; SI-NEXT:    v_mov_b32_e32 v15, v60
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[15:16], v[60:61], 16
; SI-NEXT:    v_mov_b32_e32 v16, v33
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[16:17], v[33:34], 16
; SI-NEXT:    v_mov_b32_e32 v17, v57
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[17:18], v[57:58], 16
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v47
; SI-NEXT:    v_lshr_b64 v[18:19], v[59:60], 16
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v20
; SI-NEXT:    v_lshr_b64 v[19:20], v[56:57], 16
; SI-NEXT:    v_mov_b32_e32 v20, v38
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[20:21], v[38:39], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; SI-NEXT:    v_mov_b32_e32 v39, v62
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v21
; SI-NEXT:    v_mov_b32_e32 v21, v22
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[21:22], v[22:23], 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v55
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[37:38], 16
; SI-NEXT:    v_lshr_b64 v[23:24], v[46:47], 16
; SI-NEXT:    v_lshr_b64 v[24:25], v[48:49], 16
; SI-NEXT:    v_lshr_b64 v[25:26], v[50:51], 16
; SI-NEXT:    v_lshr_b64 v[26:27], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[27:28], v[42:43], 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_lshr_b64 v[61:62], v[52:53], 16
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v45, v44
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v47, v36
; SI-NEXT:    v_mov_b32_e32 v36, v37
; SI-NEXT:    v_mov_b32_e32 v49, v50
; SI-NEXT:    v_mov_b32_e32 v51, v33
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_lshr_b64 v[29:30], v[40:41], 16
; SI-NEXT:    v_lshr_b64 v[30:31], v[54:55], 16
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v31, v61
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    s_branch .LBB79_3
; SI-NEXT:  .LBB79_2:
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v32, v59
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v58, v56
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:  .LBB79_3: ; %Flow
; SI-NEXT:    v_mov_b32_e32 v37, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v33, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1) expcnt(3)
; SI-NEXT:    v_mov_b32_e32 v38, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v50, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v40, v56
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_cbranch_vccnz .LBB79_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v55
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v34, v56
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshr_b64 v[1:2], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v33
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_lshr_b64 v[32:33], v[32:33], 16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v37
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v31, v32
; SI-NEXT:  .LBB79_5: ; %end
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v16f64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v35, s36, 0
; VI-NEXT:    v_writelane_b32 v35, s37, 1
; VI-NEXT:    v_writelane_b32 v35, s38, 2
; VI-NEXT:    v_writelane_b32 v35, s39, 3
; VI-NEXT:    v_writelane_b32 v35, s48, 4
; VI-NEXT:    v_writelane_b32 v35, s49, 5
; VI-NEXT:    v_writelane_b32 v35, s50, 6
; VI-NEXT:    v_writelane_b32 v35, s51, 7
; VI-NEXT:    v_writelane_b32 v35, s52, 8
; VI-NEXT:    v_writelane_b32 v35, s53, 9
; VI-NEXT:    v_writelane_b32 v35, s54, 10
; VI-NEXT:    v_writelane_b32 v35, s55, 11
; VI-NEXT:    v_writelane_b32 v35, s64, 12
; VI-NEXT:    v_writelane_b32 v35, s65, 13
; VI-NEXT:    v_writelane_b32 v35, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v35, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB79_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB79_4
; VI-NEXT:  .LBB79_2: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x40c00000
; VI-NEXT:    s_lshl_b32 s4, s50, 16
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    s_and_b32 s4, s50, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v2, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s48, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; VI-NEXT:    v_add_f32_e32 v2, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    s_and_b32 s4, s48, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v4, v3, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_lshrrev_b64 v[14:15], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s46, 16
; VI-NEXT:    v_lshrrev_b64 v[12:13], 16, v[2:3]
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[10:11], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[8:9], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s38, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[4:5], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s4, s36, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[2:3], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_lshl_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; VI-NEXT:    v_bfe_u32 v3, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v0
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_lshrrev_b64 v[0:1], 16, v[0:1]
; VI-NEXT:    s_and_b32 s4, s66, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s66, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s64, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s64, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s62, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[30:31], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s62, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s60, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[28:29], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s60, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s58, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[26:27], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s58, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s56, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[24:25], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s56, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s54, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[22:23], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s54, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s52, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[20:21], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s52, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    s_and_b32 s4, s37, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[15:16]
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s4, v32
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v15, v5, v7, vcc
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s4, s37, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_and_b32 s4, s39, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v5, v7, vcc
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    s_lshl_b32 s4, s39, 16
; VI-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_mov_b32_e32 v1, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v7, v9, vcc
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    s_lshl_b32 s4, s41, 16
; VI-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v3
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_mov_b32_e32 v3, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v9, v11, vcc
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    s_lshl_b32 s4, s43, 16
; VI-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_mov_b32_e32 v5, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v11, v13, vcc
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    s_lshl_b32 s4, s45, 16
; VI-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v7
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[16:17], 16, v[15:16]
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_mov_b32_e32 v7, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v13, v15, vcc
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    s_lshl_b32 s4, s47, 16
; VI-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v9
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_and_b32 s4, s49, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_mov_b32_e32 v9, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v15, v17, vcc
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    v_or_b32_e32 v17, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    s_lshl_b32 s4, s49, 16
; VI-NEXT:    v_cndmask_b32_e32 v13, v15, v17, vcc
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v11
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_and_b32 s4, s51, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_mov_b32_e32 v11, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v17, v19, vcc
; VI-NEXT:    v_bfe_u32 v17, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v15
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    s_lshl_b32 s4, s51, 16
; VI-NEXT:    v_cndmask_b32_e32 v15, v17, v19, vcc
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v13
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    s_and_b32 s4, s53, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_mov_b32_e32 v13, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v19, v21, vcc
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    s_lshl_b32 s4, s53, 16
; VI-NEXT:    v_cndmask_b32_e32 v17, v19, v21, vcc
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v15
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    s_and_b32 s4, s55, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_mov_b32_e32 v15, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v21, v23, vcc
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    s_lshl_b32 s4, s55, 16
; VI-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    s_and_b32 s4, s57, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_mov_b32_e32 v17, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v23, v25, vcc
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    s_lshl_b32 s4, s57, 16
; VI-NEXT:    v_cndmask_b32_e32 v21, v23, v25, vcc
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    s_and_b32 s4, s59, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_mov_b32_e32 v19, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v25, v27, vcc
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    s_lshl_b32 s4, s59, 16
; VI-NEXT:    v_cndmask_b32_e32 v23, v25, v27, vcc
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v21
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    s_and_b32 s4, s61, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_mov_b32_e32 v21, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v27, v29, vcc
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    s_lshl_b32 s4, s61, 16
; VI-NEXT:    v_cndmask_b32_e32 v25, v27, v29, vcc
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    s_and_b32 s4, s63, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_mov_b32_e32 v23, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v29, v31, vcc
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    s_lshl_b32 s4, s63, 16
; VI-NEXT:    v_cndmask_b32_e32 v27, v29, v31, vcc
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    s_and_b32 s4, s65, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v25, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v31, v33, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    v_mov_b32_e32 v27, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    s_lshl_b32 s4, s65, 16
; VI-NEXT:    v_cndmask_b32_e32 v29, v31, v33, vcc
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    s_and_b32 s4, s67, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_mov_b32_e32 v29, v33
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    s_lshl_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v32, v31, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_add_u32_e32 v32, vcc, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x7fff, v32
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v32, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; VI-NEXT:    v_lshrrev_b64 v[31:32], 16, v[31:32]
; VI-NEXT:    s_branch .LBB79_5
; VI-NEXT:  .LBB79_3:
; VI-NEXT:    s_branch .LBB79_2
; VI-NEXT:  .LBB79_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB79_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v35, 15
; VI-NEXT:    v_readlane_b32 s66, v35, 14
; VI-NEXT:    v_readlane_b32 s65, v35, 13
; VI-NEXT:    v_readlane_b32 s64, v35, 12
; VI-NEXT:    v_readlane_b32 s55, v35, 11
; VI-NEXT:    v_readlane_b32 s54, v35, 10
; VI-NEXT:    v_readlane_b32 s53, v35, 9
; VI-NEXT:    v_readlane_b32 s52, v35, 8
; VI-NEXT:    v_readlane_b32 s51, v35, 7
; VI-NEXT:    v_readlane_b32 s50, v35, 6
; VI-NEXT:    v_readlane_b32 s49, v35, 5
; VI-NEXT:    v_readlane_b32 s48, v35, 4
; VI-NEXT:    v_readlane_b32 s39, v35, 3
; VI-NEXT:    v_readlane_b32 s38, v35, 2
; VI-NEXT:    v_readlane_b32 s37, v35, 1
; VI-NEXT:    v_readlane_b32 s36, v35, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v16f64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v36, s36, 0
; GFX9-NEXT:    v_writelane_b32 v36, s37, 1
; GFX9-NEXT:    v_writelane_b32 v36, s38, 2
; GFX9-NEXT:    v_writelane_b32 v36, s39, 3
; GFX9-NEXT:    v_writelane_b32 v36, s48, 4
; GFX9-NEXT:    v_writelane_b32 v36, s49, 5
; GFX9-NEXT:    v_writelane_b32 v36, s50, 6
; GFX9-NEXT:    v_writelane_b32 v36, s51, 7
; GFX9-NEXT:    v_writelane_b32 v36, s52, 8
; GFX9-NEXT:    v_writelane_b32 v36, s53, 9
; GFX9-NEXT:    v_writelane_b32 v36, s54, 10
; GFX9-NEXT:    v_writelane_b32 v36, s55, 11
; GFX9-NEXT:    v_writelane_b32 v36, s64, 12
; GFX9-NEXT:    v_writelane_b32 v36, s65, 13
; GFX9-NEXT:    v_writelane_b32 v36, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v36, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB79_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB79_4
; GFX9-NEXT:  .LBB79_2: ; %cmp.true
; GFX9-NEXT:    s_and_b32 s4, s51, 0xffff0000
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x40c00000
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s51, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    s_and_b32 s4, s50, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_add_f32_e32 v2, s4, v16
; GFX9-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v3, v3, v2
; GFX9-NEXT:    v_add_u32_e32 v3, 0x7fff, v3
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    s_lshl_b32 s4, s50, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT:    v_add_f32_e32 v3, s4, v16
; GFX9-NEXT:    v_mov_b32_e32 v32, 0xffff
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    s_and_b32 s4, s49, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_lshl_or_b32 v15, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_and_b32_sdwa v3, v32, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_lshl_or_b32 v14, v2, 16, v3
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s49, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s48, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s48, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s47, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s47, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s46, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s46, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s45, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s45, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s44, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s44, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s43, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s43, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s42, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s42, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s41, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s41, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s40, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s40, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s39, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s39, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s38, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s38, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v17, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s37, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s37, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v17, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v1
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v17, v18, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s36, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v17, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v17, v17, v0
; GFX9-NEXT:    v_add_u32_e32 v17, 0x7fff, v17
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s36, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v17, v18, vcc
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v17, v32, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s67, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s67, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s66, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v31, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s66, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s65, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v30, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s65, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s64, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v29, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s64, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s63, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v28, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s63, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s62, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v27, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s62, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s61, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v26, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s61, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s60, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s60, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s59, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v24, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s59, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s58, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v23, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s58, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s57, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v22, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s57, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s56, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v21, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s56, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s55, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v20, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s55, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v33, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s54, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v19, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s54, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v33, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v18
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s53, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v17
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s53, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v33, v32, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s52, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v33
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    s_lshl_b32 s4, s52, 16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v16, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v16, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v34, v35, vcc
; GFX9-NEXT:    v_and_b32_sdwa v16, v32, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v16, v32, 16, v16
; GFX9-NEXT:    s_branch .LBB79_5
; GFX9-NEXT:  .LBB79_3:
; GFX9-NEXT:    s_branch .LBB79_2
; GFX9-NEXT:  .LBB79_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB79_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v36, 15
; GFX9-NEXT:    v_readlane_b32 s66, v36, 14
; GFX9-NEXT:    v_readlane_b32 s65, v36, 13
; GFX9-NEXT:    v_readlane_b32 s64, v36, 12
; GFX9-NEXT:    v_readlane_b32 s55, v36, 11
; GFX9-NEXT:    v_readlane_b32 s54, v36, 10
; GFX9-NEXT:    v_readlane_b32 s53, v36, 9
; GFX9-NEXT:    v_readlane_b32 s52, v36, 8
; GFX9-NEXT:    v_readlane_b32 s51, v36, 7
; GFX9-NEXT:    v_readlane_b32 s50, v36, 6
; GFX9-NEXT:    v_readlane_b32 s49, v36, 5
; GFX9-NEXT:    v_readlane_b32 s48, v36, 4
; GFX9-NEXT:    v_readlane_b32 s39, v36, 3
; GFX9-NEXT:    v_readlane_b32 s38, v36, 2
; GFX9-NEXT:    v_readlane_b32 s37, v36, 1
; GFX9-NEXT:    v_readlane_b32 s36, v36, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v16f64_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s36, s0
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-TRUE16-NEXT:    s_mov_b32 s47, s23
; GFX11-TRUE16-NEXT:    s_mov_b32 s46, s22
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-TRUE16-NEXT:    s_mov_b32 s45, s21
; GFX11-TRUE16-NEXT:    s_mov_b32 s44, s20
; GFX11-TRUE16-NEXT:    s_mov_b32 s43, s19
; GFX11-TRUE16-NEXT:    s_mov_b32 s42, s18
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-TRUE16-NEXT:    s_mov_b32 s48, s24
; GFX11-TRUE16-NEXT:    s_mov_b32 s41, s17
; GFX11-TRUE16-NEXT:    s_mov_b32 s40, s16
; GFX11-TRUE16-NEXT:    s_mov_b32 s39, s3
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-TRUE16-NEXT:    s_mov_b32 s49, s25
; GFX11-TRUE16-NEXT:    s_mov_b32 s38, s2
; GFX11-TRUE16-NEXT:    s_mov_b32 s37, s1
; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-TRUE16-NEXT:    s_mov_b32 s50, s26
; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-TRUE16-NEXT:    s_mov_b32 s51, s27
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-TRUE16-NEXT:    s_mov_b32 s52, s28
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-TRUE16-NEXT:    s_mov_b32 s53, s29
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB79_3
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB79_4
; GFX11-TRUE16-NEXT:  .LBB79_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s51, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s50, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s57, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v0.l
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v2.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v4, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v7, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v2.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v8, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v16, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v7 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v18, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v0.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v17 :: v_dual_add_nc_u32 v0, v3, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v16, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v0, v16 :: v_dual_add_nc_u32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v2.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v17, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v21, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v3, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s37, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v0.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v1, v20 :: v_dual_add_nc_u32 v0, v16, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s36, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v22, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v16.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v1, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v19, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s67, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v24, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s66, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s65, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v20, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v20 :: v_dual_add_nc_u32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s64, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s63, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s62, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v25, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v25
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s61, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v24, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v16.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s60, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v16.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v23 :: v_dual_add_nc_u32 v16, v19, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v23, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v18.l
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v21, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s58, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v33, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s57, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v19, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v33 :: v_dual_add_nc_u32 v19, v19, v20
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v17.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v21, v32
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v16, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v33, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v20, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v17, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s55, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v20, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v16.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v35, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v18.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s52, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v19, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v33, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, v32, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v19, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, v35, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v34, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v17.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v36, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, v16, v34
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v32, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v32, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v36, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v16, v35 :: v_dual_add_nc_u32 v35, v37, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, v38, v33
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v37
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, v34, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v37, v38, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x7fff, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v17.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v35, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v32
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v34, v48, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v16.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v34
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v33.l
; GFX11-TRUE16-NEXT:    s_branch .LBB79_5
; GFX11-TRUE16-NEXT:  .LBB79_3:
; GFX11-TRUE16-NEXT:    s_branch .LBB79_2
; GFX11-TRUE16-NEXT:  .LBB79_4:
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-TRUE16-NEXT:  .LBB79_5: ; %end
; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v16f64_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s36, s0
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-FAKE16-NEXT:    s_mov_b32 s47, s23
; GFX11-FAKE16-NEXT:    s_mov_b32 s46, s22
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-FAKE16-NEXT:    s_mov_b32 s45, s21
; GFX11-FAKE16-NEXT:    s_mov_b32 s44, s20
; GFX11-FAKE16-NEXT:    s_mov_b32 s43, s19
; GFX11-FAKE16-NEXT:    s_mov_b32 s42, s18
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-FAKE16-NEXT:    s_mov_b32 s48, s24
; GFX11-FAKE16-NEXT:    s_mov_b32 s41, s17
; GFX11-FAKE16-NEXT:    s_mov_b32 s40, s16
; GFX11-FAKE16-NEXT:    s_mov_b32 s39, s3
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-FAKE16-NEXT:    s_mov_b32 s49, s25
; GFX11-FAKE16-NEXT:    s_mov_b32 s38, s2
; GFX11-FAKE16-NEXT:    s_mov_b32 s37, s1
; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-FAKE16-NEXT:    s_mov_b32 s50, s26
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-FAKE16-NEXT:    s_mov_b32 s51, s27
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s52, 8
; GFX11-FAKE16-NEXT:    s_mov_b32 s52, s28
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s53, 9
; GFX11-FAKE16-NEXT:    s_mov_b32 s53, s29
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s54, 10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s55, 11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s64, 12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s65, 13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s66, 14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s67, 15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB79_3
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB79_4
; GFX11-FAKE16-NEXT:  .LBB79_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s51, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s2, s50, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s49, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s3, s49, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s48, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v2, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v9, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v10, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v4, v7 :: v_dual_add_nc_u32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v7, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v5, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s47, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s46, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s45, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s44, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s43, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v8, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s42, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v7, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s41, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s40, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v17 :: v_dual_add_nc_u32 v5, v7, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s39, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s38, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v17, v18
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s37, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v4, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s67, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s67, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v2, v22 :: v_dual_add_nc_u32 v17, v17, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v22, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s66, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s66, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s65, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s65, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v0, 16, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v22, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s64, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s64, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s63, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s63, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s62, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s62, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s61, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s61, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v24, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s60, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v23, v24
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s60, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v16, 16, v17
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s59, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v24, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v20, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s58, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s59, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v20, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s58, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v32, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, v22, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v19, v19, v21 :: v_dual_and_b32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s57, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v18, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v21, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s57, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v17
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v22, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v23, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v33, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v35, v23
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s56, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s56, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v20, v32, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v21, v33, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v16, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v33, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v22, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s55, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s55, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s54, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v32, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v18, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s54, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, v34, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s53, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v32, v33, v34 :: v_dual_add_nc_u32 v33, v35, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s53, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v37, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s52, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v33, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v33, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s52, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v38, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, v35, v37
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, v36, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, v39, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x7fff, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x7fff, v36
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, v48, v38
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x7fff, v39
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x7fff, v48
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v39, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v36, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v37, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v17, 16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v34, 16, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v32, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v36, 16, v33
; GFX11-FAKE16-NEXT:    s_branch .LBB79_5
; GFX11-FAKE16-NEXT:  .LBB79_3:
; GFX11-FAKE16-NEXT:    s_branch .LBB79_2
; GFX11-FAKE16-NEXT:  .LBB79_4:
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-FAKE16-NEXT:  .LBB79_5: ; %end
; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v40, 15
; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v40, 14
; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v40, 13
; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v40, 12
; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v40, 11
; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v40, 10
; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v40, 9
; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v40, 8
; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define <64 x half> @bitcast_v16f64_to_v64f16(<16 x double> %a, i32 %b) {
; SI-LABEL: bitcast_v16f64_to_v64f16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB80_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v40, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v43, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v46, v3, v2, 16
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_alignbit_b32 v57, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB80_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB80_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; SI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v40, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v43, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v46, v3, v2, 16
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_alignbit_b32 v57, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB80_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v57
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v46
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v43
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; SI-NEXT:    v_or_b32_e32 v0, v0, v57
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v46
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v43
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v61
; SI-NEXT:    v_or_b32_e32 v6, v6, v40
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v60
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_or_b32_e32 v1, v1, v57
; SI-NEXT:    v_or_b32_e32 v3, v3, v46
; SI-NEXT:    v_or_b32_e32 v5, v5, v43
; SI-NEXT:    v_or_b32_e32 v7, v7, v40
; SI-NEXT:    v_or_b32_e32 v8, v8, v52
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v56
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v47
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v45
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v44
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v42
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v41
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v55
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v54
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v53
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v9, v9, v52
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v64f16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB80_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; VI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; VI-NEXT:  .LBB80_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v64f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB80_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX9-NEXT:  .LBB80_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v64f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB80_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX11-NEXT:  .LBB80_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define inreg <64 x half> @bitcast_v16f64_to_v64f16_scalar(<16 x double> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16f64_to_v64f16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_readfirstlane_b32 s45, v17
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    v_readfirstlane_b32 s43, v15
; SI-NEXT:    v_readfirstlane_b32 s42, v14
; SI-NEXT:    v_readfirstlane_b32 s41, v13
; SI-NEXT:    v_readfirstlane_b32 s40, v12
; SI-NEXT:    v_readfirstlane_b32 s15, v11
; SI-NEXT:    v_readfirstlane_b32 s14, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s11, v7
; SI-NEXT:    v_readfirstlane_b32 s10, v6
; SI-NEXT:    v_readfirstlane_b32 s9, v5
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_readfirstlane_b32 s7, v3
; SI-NEXT:    v_readfirstlane_b32 s6, v2
; SI-NEXT:    v_readfirstlane_b32 s5, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s4, v0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB81_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s69, s45, 16
; SI-NEXT:    s_lshr_b32 s68, s43, 16
; SI-NEXT:    s_lshr_b32 s67, s41, 16
; SI-NEXT:    s_lshr_b32 s66, s15, 16
; SI-NEXT:    s_lshr_b32 s65, s13, 16
; SI-NEXT:    s_lshr_b32 s64, s11, 16
; SI-NEXT:    s_lshr_b32 s55, s9, 16
; SI-NEXT:    s_lshr_b32 s54, s7, 16
; SI-NEXT:    s_lshr_b32 s53, s5, 16
; SI-NEXT:    s_lshr_b32 s52, s29, 16
; SI-NEXT:    s_lshr_b32 s51, s27, 16
; SI-NEXT:    s_lshr_b32 s50, s25, 16
; SI-NEXT:    s_lshr_b32 s49, s23, 16
; SI-NEXT:    s_lshr_b32 s48, s21, 16
; SI-NEXT:    s_lshr_b32 s39, s19, 16
; SI-NEXT:    s_lshr_b32 s38, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB81_4
; SI-NEXT:  .LBB81_2: ; %cmp.true
; SI-NEXT:    v_add_f64 v[28:29], s[42:43], 1.0
; SI-NEXT:    v_add_f64 v[26:27], s[40:41], 1.0
; SI-NEXT:    v_add_f64 v[30:31], s[44:45], 1.0
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v29
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v27
; SI-NEXT:    v_add_f64 v[16:17], s[6:7], 1.0
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[32:33], v[30:31], 16
; SI-NEXT:    v_add_f64 v[24:25], s[14:15], 1.0
; SI-NEXT:    v_add_f64 v[14:15], s[4:5], 1.0
; SI-NEXT:    v_lshr_b64 v[33:34], v[28:29], 16
; SI-NEXT:    v_add_f64 v[22:23], s[12:13], 1.0
; SI-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; SI-NEXT:    v_lshr_b64 v[34:35], v[26:27], 16
; SI-NEXT:    v_lshr_b64 v[48:49], v[16:17], 16
; SI-NEXT:    v_add_f64 v[20:21], s[10:11], 1.0
; SI-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; SI-NEXT:    v_lshr_b64 v[35:36], v[24:25], 16
; SI-NEXT:    v_lshr_b64 v[49:50], v[14:15], 16
; SI-NEXT:    v_add_f64 v[18:19], s[8:9], 1.0
; SI-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; SI-NEXT:    v_lshr_b64 v[36:37], v[22:23], 16
; SI-NEXT:    v_lshr_b64 v[50:51], v[12:13], 16
; SI-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; SI-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; SI-NEXT:    v_lshr_b64 v[37:38], v[20:21], 16
; SI-NEXT:    v_lshr_b64 v[51:52], v[10:11], 16
; SI-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; SI-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; SI-NEXT:    v_lshr_b64 v[38:39], v[18:19], 16
; SI-NEXT:    v_lshr_b64 v[52:53], v[8:9], 16
; SI-NEXT:    v_lshr_b64 v[53:54], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[39:40], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[54:55], v[4:5], 16
; SI-NEXT:    v_lshr_b64 v[40:41], v[0:1], 16
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v1
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v31
; SI-NEXT:    s_branch .LBB81_5
; SI-NEXT:  .LBB81_3:
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    s_branch .LBB81_2
; SI-NEXT:  .LBB81_4:
; SI-NEXT:    v_mov_b32_e32 v32, s68
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v32, s67
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v31, s45
; SI-NEXT:    v_mov_b32_e32 v29, s43
; SI-NEXT:    v_mov_b32_e32 v27, s41
; SI-NEXT:    v_mov_b32_e32 v25, s15
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v21, s11
; SI-NEXT:    v_mov_b32_e32 v19, s9
; SI-NEXT:    v_mov_b32_e32 v17, s7
; SI-NEXT:    v_mov_b32_e32 v15, s5
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v14, s4
; SI-NEXT:    v_mov_b32_e32 v16, s6
; SI-NEXT:    v_mov_b32_e32 v18, s8
; SI-NEXT:    v_mov_b32_e32 v20, s10
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v24, s14
; SI-NEXT:    v_mov_b32_e32 v26, s40
; SI-NEXT:    v_mov_b32_e32 v28, s42
; SI-NEXT:    v_mov_b32_e32 v30, s44
; SI-NEXT:    v_mov_b32_e32 v41, s69
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v45, s66
; SI-NEXT:    v_mov_b32_e32 v46, s65
; SI-NEXT:    v_mov_b32_e32 v47, s64
; SI-NEXT:    v_mov_b32_e32 v56, s55
; SI-NEXT:    v_mov_b32_e32 v57, s54
; SI-NEXT:    v_mov_b32_e32 v58, s53
; SI-NEXT:    v_mov_b32_e32 v59, s52
; SI-NEXT:    v_mov_b32_e32 v60, s51
; SI-NEXT:    v_mov_b32_e32 v61, s50
; SI-NEXT:    v_mov_b32_e32 v62, s49
; SI-NEXT:    v_mov_b32_e32 v42, s48
; SI-NEXT:    v_mov_b32_e32 v43, s39
; SI-NEXT:    v_mov_b32_e32 v44, s38
; SI-NEXT:    v_mov_b32_e32 v40, s36
; SI-NEXT:    v_mov_b32_e32 v39, s34
; SI-NEXT:    v_mov_b32_e32 v54, s30
; SI-NEXT:    v_mov_b32_e32 v53, s94
; SI-NEXT:    v_mov_b32_e32 v52, s92
; SI-NEXT:    v_mov_b32_e32 v51, s90
; SI-NEXT:    v_mov_b32_e32 v50, s88
; SI-NEXT:    v_mov_b32_e32 v49, s78
; SI-NEXT:    v_mov_b32_e32 v48, s74
; SI-NEXT:    v_mov_b32_e32 v38, s72
; SI-NEXT:    v_mov_b32_e32 v37, s62
; SI-NEXT:    v_mov_b32_e32 v36, s58
; SI-NEXT:    v_mov_b32_e32 v35, s56
; SI-NEXT:    v_mov_b32_e32 v34, s76
; SI-NEXT:    v_mov_b32_e32 v33, s60
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v32, s46
; SI-NEXT:  .LBB81_5: ; %end
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v39
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v43
; SI-NEXT:    v_or_b32_e32 v3, v3, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v54
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v39
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v42
; SI-NEXT:    v_or_b32_e32 v5, v5, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v53
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v39
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v62
; SI-NEXT:    v_or_b32_e32 v7, v7, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v52
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v39
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v61
; SI-NEXT:    v_or_b32_e32 v9, v9, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v51
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v39
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v60
; SI-NEXT:    v_or_b32_e32 v11, v11, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v39
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v59
; SI-NEXT:    v_or_b32_e32 v13, v13, v39
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v49
; SI-NEXT:    v_or_b32_e32 v14, v14, v39
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v58
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v40
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v15, v15, v39
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v48
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v0, v0, v55
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v44
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v57
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v56
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v47
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v46
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v45
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v41
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v1, v1, v55
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v64f16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s36, 2
; VI-NEXT:    v_writelane_b32 v32, s37, 3
; VI-NEXT:    v_writelane_b32 v32, s38, 4
; VI-NEXT:    v_writelane_b32 v32, s39, 5
; VI-NEXT:    v_writelane_b32 v32, s48, 6
; VI-NEXT:    v_writelane_b32 v32, s49, 7
; VI-NEXT:    v_writelane_b32 v32, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB81_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB81_4
; VI-NEXT:  .LBB81_2: ; %cmp.true
; VI-NEXT:    v_add_f64 v[14:15], s[30:31], 1.0
; VI-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; VI-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; VI-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; VI-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; VI-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; VI-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; VI-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; VI-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; VI-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; VI-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; VI-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; VI-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; VI-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; VI-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; VI-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; VI-NEXT:    s_branch .LBB81_5
; VI-NEXT:  .LBB81_3:
; VI-NEXT:    s_branch .LBB81_2
; VI-NEXT:  .LBB81_4:
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:  .LBB81_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v32, 9
; VI-NEXT:    v_readlane_b32 s50, v32, 8
; VI-NEXT:    v_readlane_b32 s49, v32, 7
; VI-NEXT:    v_readlane_b32 s48, v32, 6
; VI-NEXT:    v_readlane_b32 s39, v32, 5
; VI-NEXT:    v_readlane_b32 s38, v32, 4
; VI-NEXT:    v_readlane_b32 s37, v32, 3
; VI-NEXT:    v_readlane_b32 s36, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v64f16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB81_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB81_4
; GFX9-NEXT:  .LBB81_2: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[14:15], s[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; GFX9-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; GFX9-NEXT:    s_branch .LBB81_5
; GFX9-NEXT:  .LBB81_3:
; GFX9-NEXT:    s_branch .LBB81_2
; GFX9-NEXT:  .LBB81_4:
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:  .LBB81_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v64f16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB81_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB81_4
; GFX11-NEXT:  .LBB81_2: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[14:15], s[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], s[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], s[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], s[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], s[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], s[16:17], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], s[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], s[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; GFX11-NEXT:    s_branch .LBB81_5
; GFX11-NEXT:  .LBB81_3:
; GFX11-NEXT:    s_branch .LBB81_2
; GFX11-NEXT:  .LBB81_4:
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:  .LBB81_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define <16 x double> @bitcast_v64f16_to_v16f64(<64 x half> %a, i32 %b) {
; SI-LABEL: bitcast_v64f16_to_v16f64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v30
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v29
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v27
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v25
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v24
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v23
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v22
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v21
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v20
; SI-NEXT:    v_mov_b32_e32 v52, v18
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
; SI-NEXT:    v_mov_b32_e32 v53, v17
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v52
; SI-NEXT:    v_mov_b32_e32 v54, v16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v53
; SI-NEXT:    v_mov_b32_e32 v55, v15
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v54
; SI-NEXT:    v_mov_b32_e32 v40, v14
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v55
; SI-NEXT:    v_mov_b32_e32 v41, v13
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v40
; SI-NEXT:    v_mov_b32_e32 v42, v12
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v28
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v42
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v61
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v62
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB82_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v63
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v61
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v50
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v32
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v36
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v49
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v33
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v48
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v35
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v39
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v34
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v38
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v37
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v42
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; kill: killed $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v41
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v40
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v55
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v53
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v52
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB82_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB82_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, v63
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v50
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v62
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v61
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_or_b32_e32 v1, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v59
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v58
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v36
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v57
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v56
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v49
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v47
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v46
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v33
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v45
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v44
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v48
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v43
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v42
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v35
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v41
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v40
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v39
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v55
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v54
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v34
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v53
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v19, v52
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v38
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v37
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v51
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:  .LBB82_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v16f64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB82_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x200
; VI-NEXT:    v_add_f16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v15, 0x200, v15
; VI-NEXT:    v_or_b32_e32 v15, v15, v33
; VI-NEXT:    v_add_f16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v14, 0x200, v14
; VI-NEXT:    v_or_b32_e32 v14, v14, v33
; VI-NEXT:    v_add_f16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v13, 0x200, v13
; VI-NEXT:    v_or_b32_e32 v13, v13, v33
; VI-NEXT:    v_add_f16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v12, 0x200, v12
; VI-NEXT:    v_or_b32_e32 v12, v12, v33
; VI-NEXT:    v_add_f16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v11, 0x200, v11
; VI-NEXT:    v_or_b32_e32 v11, v11, v33
; VI-NEXT:    v_add_f16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v10, 0x200, v10
; VI-NEXT:    v_or_b32_e32 v10, v10, v33
; VI-NEXT:    v_add_f16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v9, 0x200, v9
; VI-NEXT:    v_or_b32_e32 v9, v9, v33
; VI-NEXT:    v_add_f16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v8, 0x200, v8
; VI-NEXT:    v_or_b32_e32 v8, v8, v33
; VI-NEXT:    v_add_f16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v7, 0x200, v7
; VI-NEXT:    v_or_b32_e32 v7, v7, v33
; VI-NEXT:    v_add_f16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v6, 0x200, v6
; VI-NEXT:    v_or_b32_e32 v6, v6, v33
; VI-NEXT:    v_add_f16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v5, 0x200, v5
; VI-NEXT:    v_or_b32_e32 v5, v5, v33
; VI-NEXT:    v_add_f16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v4, 0x200, v4
; VI-NEXT:    v_or_b32_e32 v4, v4, v33
; VI-NEXT:    v_add_f16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, 0x200, v3
; VI-NEXT:    v_or_b32_e32 v3, v3, v33
; VI-NEXT:    v_add_f16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v2, 0x200, v2
; VI-NEXT:    v_or_b32_e32 v2, v2, v33
; VI-NEXT:    v_add_f16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, 0x200, v1
; VI-NEXT:    v_or_b32_e32 v1, v1, v33
; VI-NEXT:    v_add_f16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, 0x200, v0
; VI-NEXT:    v_or_b32_e32 v0, v0, v33
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v31, 0x200, v31
; VI-NEXT:    v_or_b32_e32 v31, v31, v33
; VI-NEXT:    v_add_f16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v30, 0x200, v30
; VI-NEXT:    v_or_b32_e32 v30, v30, v33
; VI-NEXT:    v_add_f16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v29, 0x200, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    v_add_f16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v28, 0x200, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v33
; VI-NEXT:    v_add_f16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v27, 0x200, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v33
; VI-NEXT:    v_add_f16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v26, 0x200, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v33
; VI-NEXT:    v_add_f16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v25, 0x200, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v33
; VI-NEXT:    v_add_f16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v24, 0x200, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v33
; VI-NEXT:    v_add_f16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v23, 0x200, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v33
; VI-NEXT:    v_add_f16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v22, 0x200, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v33
; VI-NEXT:    v_add_f16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v21, 0x200, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v33
; VI-NEXT:    v_add_f16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v20, 0x200, v20
; VI-NEXT:    v_or_b32_e32 v20, v20, v33
; VI-NEXT:    v_add_f16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v19, 0x200, v19
; VI-NEXT:    v_or_b32_e32 v19, v19, v33
; VI-NEXT:    v_add_f16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, 0x200, v18
; VI-NEXT:    v_or_b32_e32 v18, v18, v33
; VI-NEXT:    v_add_f16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, 0x200, v17
; VI-NEXT:    v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, 0x200, v16
; VI-NEXT:    v_or_b32_e32 v17, v17, v33
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB82_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v16f64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB82_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_movk_i32 s6, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB82_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v16f64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB82_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-NEXT:  .LBB82_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define inreg <16 x double> @bitcast_v64f16_to_v16f64_scalar(<64 x half> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64f16_to_v16f64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v34, s30, 0
; SI-NEXT:    v_writelane_b32 v34, s31, 1
; SI-NEXT:    v_writelane_b32 v34, s34, 2
; SI-NEXT:    v_writelane_b32 v34, s35, 3
; SI-NEXT:    v_writelane_b32 v34, s36, 4
; SI-NEXT:    v_writelane_b32 v34, s37, 5
; SI-NEXT:    v_writelane_b32 v34, s38, 6
; SI-NEXT:    v_writelane_b32 v34, s39, 7
; SI-NEXT:    v_writelane_b32 v34, s48, 8
; SI-NEXT:    v_writelane_b32 v34, s49, 9
; SI-NEXT:    v_writelane_b32 v34, s50, 10
; SI-NEXT:    v_writelane_b32 v34, s51, 11
; SI-NEXT:    v_writelane_b32 v34, s52, 12
; SI-NEXT:    v_writelane_b32 v34, s53, 13
; SI-NEXT:    v_writelane_b32 v34, s54, 14
; SI-NEXT:    v_writelane_b32 v34, s55, 15
; SI-NEXT:    v_writelane_b32 v34, s64, 16
; SI-NEXT:    v_writelane_b32 v34, s65, 17
; SI-NEXT:    v_writelane_b32 v34, s66, 18
; SI-NEXT:    v_writelane_b32 v34, s67, 19
; SI-NEXT:    v_writelane_b32 v34, s68, 20
; SI-NEXT:    v_writelane_b32 v34, s69, 21
; SI-NEXT:    v_writelane_b32 v34, s70, 22
; SI-NEXT:    v_writelane_b32 v34, s71, 23
; SI-NEXT:    v_writelane_b32 v34, s80, 24
; SI-NEXT:    v_writelane_b32 v34, s81, 25
; SI-NEXT:    v_writelane_b32 v34, s82, 26
; SI-NEXT:    v_writelane_b32 v34, s83, 27
; SI-NEXT:    v_writelane_b32 v34, s84, 28
; SI-NEXT:    v_writelane_b32 v34, s85, 29
; SI-NEXT:    v_writelane_b32 v34, s86, 30
; SI-NEXT:    v_writelane_b32 v34, s87, 31
; SI-NEXT:    v_writelane_b32 v34, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s6, v17
; SI-NEXT:    v_writelane_b32 v34, s97, 33
; SI-NEXT:    s_lshr_b32 vcc_lo, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s8, v16
; SI-NEXT:    ; implicit-def: $vgpr35 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v34, s98, 34
; SI-NEXT:    s_lshr_b32 vcc_hi, s8, 16
; SI-NEXT:    v_readfirstlane_b32 s10, v15
; SI-NEXT:    v_readfirstlane_b32 s12, v14
; SI-NEXT:    v_readfirstlane_b32 s14, v13
; SI-NEXT:    v_readfirstlane_b32 s72, v12
; SI-NEXT:    v_readfirstlane_b32 s74, v11
; SI-NEXT:    v_readfirstlane_b32 s76, v10
; SI-NEXT:    v_readfirstlane_b32 s78, v9
; SI-NEXT:    v_readfirstlane_b32 s88, v8
; SI-NEXT:    v_readfirstlane_b32 s90, v7
; SI-NEXT:    v_readfirstlane_b32 s92, v6
; SI-NEXT:    v_readfirstlane_b32 s94, v5
; SI-NEXT:    v_readfirstlane_b32 s30, v4
; SI-NEXT:    v_readfirstlane_b32 s34, v3
; SI-NEXT:    v_readfirstlane_b32 s68, v2
; SI-NEXT:    v_readfirstlane_b32 s80, v1
; SI-NEXT:    v_readfirstlane_b32 s83, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v35, vcc_lo, 0
; SI-NEXT:    v_writelane_b32 v34, s99, 35
; SI-NEXT:    s_lshr_b32 s69, s29, 16
; SI-NEXT:    s_lshr_b32 s71, s28, 16
; SI-NEXT:    s_lshr_b32 s82, s27, 16
; SI-NEXT:    s_lshr_b32 s85, s26, 16
; SI-NEXT:    s_lshr_b32 s86, s25, 16
; SI-NEXT:    s_lshr_b32 s87, s24, 16
; SI-NEXT:    s_lshr_b32 s96, s23, 16
; SI-NEXT:    s_lshr_b32 s97, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s7, s19, 16
; SI-NEXT:    s_lshr_b32 s9, s18, 16
; SI-NEXT:    s_lshr_b32 s11, s17, 16
; SI-NEXT:    s_lshr_b32 s13, s16, 16
; SI-NEXT:    s_lshr_b32 s65, s10, 16
; SI-NEXT:    s_lshr_b32 s64, s12, 16
; SI-NEXT:    s_lshr_b32 s15, s14, 16
; SI-NEXT:    s_lshr_b32 s73, s72, 16
; SI-NEXT:    s_lshr_b32 s75, s74, 16
; SI-NEXT:    s_lshr_b32 s77, s76, 16
; SI-NEXT:    s_lshr_b32 s79, s78, 16
; SI-NEXT:    s_lshr_b32 s89, s88, 16
; SI-NEXT:    s_lshr_b32 s91, s90, 16
; SI-NEXT:    s_lshr_b32 s93, s92, 16
; SI-NEXT:    s_lshr_b32 s95, s94, 16
; SI-NEXT:    s_lshr_b32 s31, s30, 16
; SI-NEXT:    s_lshr_b32 s35, s34, 16
; SI-NEXT:    s_lshr_b32 s70, s68, 16
; SI-NEXT:    s_lshr_b32 s81, s80, 16
; SI-NEXT:    s_lshr_b32 s84, s83, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v35, vcc_hi, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v35, s65, 2
; SI-NEXT:    v_writelane_b32 v35, s64, 3
; SI-NEXT:    s_cbranch_scc0 .LBB83_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s13, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s9, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s7, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s11, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s97, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s86, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s34, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s30, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s31, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s92, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s90, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s91, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s88, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s89, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s78, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s79, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s76, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s77, 16
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s74, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s75, 16
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s72, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s73, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s15, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s64, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s65, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB83_4
; SI-NEXT:  .LBB83_2: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s13
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s16
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s11
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s17
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s18
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s9
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v1, v3, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s7
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s19
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s99
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s20
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s98
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s21
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s97
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v7, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s22
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s96
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s23
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s87
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s24
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_or_b32_e32 v7, v9, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s86
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v10, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s25
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s85
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s26
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s82
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s27
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_or_b32_e32 v10, v12, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s71
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v13, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s28
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s69
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s29
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s84
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s83
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_or_b32_e32 v13, v15, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s81
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v16, s80
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s70
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s68
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v17, s35
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s34
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_or_b32_e32 v16, v18, v16
; SI-NEXT:    v_cvt_f32_f16_e32 v18, s31
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v19, s30
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s95
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s94
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v20, s93
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s92
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    v_cvt_f32_f16_e32 v21, s91
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v22, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v22, s90
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s89
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s88
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v23, s79
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s78
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_or_b32_e32 v22, v24, v22
; SI-NEXT:    v_cvt_f32_f16_e32 v24, s77
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v25, s76
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s75
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s74
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v26, s73
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s72
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    v_cvt_f32_f16_e32 v27, s15
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_readlane_b32 s4, v35, 3
; SI-NEXT:    v_or_b32_e32 v26, v28, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v28, s14
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_readlane_b32 s4, v35, 2
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s12
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s10
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_readlane_b32 s4, v35, 1
; SI-NEXT:    v_or_b32_e32 v28, v30, v28
; SI-NEXT:    v_cvt_f32_f16_e32 v30, s4
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_readlane_b32 s4, v35, 0
; SI-NEXT:    v_or_b32_e32 v29, v31, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s8
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s4
; SI-NEXT:    v_cvt_f32_f16_e32 v33, s6
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v32
; SI-NEXT:    v_or_b32_e32 v31, v33, v31
; SI-NEXT:    s_branch .LBB83_5
; SI-NEXT:  .LBB83_3:
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB83_2
; SI-NEXT:  .LBB83_4:
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:  .LBB83_5: ; %end
; SI-NEXT:    v_readlane_b32 s99, v34, 35
; SI-NEXT:    v_readlane_b32 s98, v34, 34
; SI-NEXT:    v_readlane_b32 s97, v34, 33
; SI-NEXT:    v_readlane_b32 s96, v34, 32
; SI-NEXT:    v_readlane_b32 s87, v34, 31
; SI-NEXT:    v_readlane_b32 s86, v34, 30
; SI-NEXT:    v_readlane_b32 s85, v34, 29
; SI-NEXT:    v_readlane_b32 s84, v34, 28
; SI-NEXT:    v_readlane_b32 s83, v34, 27
; SI-NEXT:    v_readlane_b32 s82, v34, 26
; SI-NEXT:    v_readlane_b32 s81, v34, 25
; SI-NEXT:    v_readlane_b32 s80, v34, 24
; SI-NEXT:    v_readlane_b32 s71, v34, 23
; SI-NEXT:    v_readlane_b32 s70, v34, 22
; SI-NEXT:    v_readlane_b32 s69, v34, 21
; SI-NEXT:    v_readlane_b32 s68, v34, 20
; SI-NEXT:    v_readlane_b32 s67, v34, 19
; SI-NEXT:    v_readlane_b32 s66, v34, 18
; SI-NEXT:    v_readlane_b32 s65, v34, 17
; SI-NEXT:    v_readlane_b32 s64, v34, 16
; SI-NEXT:    v_readlane_b32 s55, v34, 15
; SI-NEXT:    v_readlane_b32 s54, v34, 14
; SI-NEXT:    v_readlane_b32 s53, v34, 13
; SI-NEXT:    v_readlane_b32 s52, v34, 12
; SI-NEXT:    v_readlane_b32 s51, v34, 11
; SI-NEXT:    v_readlane_b32 s50, v34, 10
; SI-NEXT:    v_readlane_b32 s49, v34, 9
; SI-NEXT:    v_readlane_b32 s48, v34, 8
; SI-NEXT:    v_readlane_b32 s39, v34, 7
; SI-NEXT:    v_readlane_b32 s38, v34, 6
; SI-NEXT:    v_readlane_b32 s37, v34, 5
; SI-NEXT:    v_readlane_b32 s36, v34, 4
; SI-NEXT:    v_readlane_b32 s35, v34, 3
; SI-NEXT:    v_readlane_b32 s34, v34, 2
; SI-NEXT:    v_readlane_b32 s31, v34, 1
; SI-NEXT:    v_readlane_b32 s30, v34, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v16f64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v33, s36, 0
; VI-NEXT:    v_writelane_b32 v33, s37, 1
; VI-NEXT:    v_writelane_b32 v33, s38, 2
; VI-NEXT:    v_writelane_b32 v33, s39, 3
; VI-NEXT:    v_writelane_b32 v33, s48, 4
; VI-NEXT:    v_writelane_b32 v33, s49, 5
; VI-NEXT:    v_writelane_b32 v33, s50, 6
; VI-NEXT:    v_writelane_b32 v33, s51, 7
; VI-NEXT:    v_writelane_b32 v33, s52, 8
; VI-NEXT:    v_writelane_b32 v33, s53, 9
; VI-NEXT:    v_writelane_b32 v33, s54, 10
; VI-NEXT:    v_writelane_b32 v33, s55, 11
; VI-NEXT:    v_writelane_b32 v33, s64, 12
; VI-NEXT:    v_writelane_b32 v33, s65, 13
; VI-NEXT:    v_writelane_b32 v33, s66, 14
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v33, s67, 15
; VI-NEXT:    s_mov_b32 s49, s29
; VI-NEXT:    s_mov_b32 s48, s28
; VI-NEXT:    s_mov_b32 s47, s27
; VI-NEXT:    s_mov_b32 s46, s26
; VI-NEXT:    s_mov_b32 s45, s25
; VI-NEXT:    s_mov_b32 s44, s24
; VI-NEXT:    s_mov_b32 s43, s23
; VI-NEXT:    s_mov_b32 s42, s22
; VI-NEXT:    s_mov_b32 s41, s21
; VI-NEXT:    s_mov_b32 s40, s20
; VI-NEXT:    s_mov_b32 s39, s19
; VI-NEXT:    s_mov_b32 s38, s18
; VI-NEXT:    s_mov_b32 s37, s17
; VI-NEXT:    s_mov_b32 s36, s16
; VI-NEXT:    v_readfirstlane_b32 s67, v17
; VI-NEXT:    v_readfirstlane_b32 s66, v16
; VI-NEXT:    v_readfirstlane_b32 s65, v15
; VI-NEXT:    v_readfirstlane_b32 s64, v14
; VI-NEXT:    v_readfirstlane_b32 s63, v13
; VI-NEXT:    v_readfirstlane_b32 s62, v12
; VI-NEXT:    v_readfirstlane_b32 s61, v11
; VI-NEXT:    v_readfirstlane_b32 s60, v10
; VI-NEXT:    v_readfirstlane_b32 s59, v9
; VI-NEXT:    v_readfirstlane_b32 s58, v8
; VI-NEXT:    v_readfirstlane_b32 s57, v7
; VI-NEXT:    v_readfirstlane_b32 s56, v6
; VI-NEXT:    v_readfirstlane_b32 s55, v5
; VI-NEXT:    v_readfirstlane_b32 s54, v4
; VI-NEXT:    v_readfirstlane_b32 s53, v3
; VI-NEXT:    v_readfirstlane_b32 s52, v2
; VI-NEXT:    v_readfirstlane_b32 s51, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s50, v0
; VI-NEXT:    s_cbranch_scc0 .LBB83_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB83_4
; VI-NEXT:  .LBB83_2: ; %cmp.true
; VI-NEXT:    s_lshr_b32 s4, s51, 16
; VI-NEXT:    v_mov_b32_e32 v16, 0x200
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    s_lshr_b32 s4, s50, 16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s51, v16
; VI-NEXT:    v_mov_b32_e32 v2, s4
; VI-NEXT:    s_lshr_b32 s4, s49, 16
; VI-NEXT:    v_or_b32_e32 v15, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s49, v16
; VI-NEXT:    s_lshr_b32 s4, s48, 16
; VI-NEXT:    v_or_b32_e32 v13, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s48, v16
; VI-NEXT:    s_lshr_b32 s4, s47, 16
; VI-NEXT:    v_or_b32_e32 v12, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s47, v16
; VI-NEXT:    s_lshr_b32 s4, s46, 16
; VI-NEXT:    v_or_b32_e32 v11, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s46, v16
; VI-NEXT:    s_lshr_b32 s4, s45, 16
; VI-NEXT:    v_or_b32_e32 v10, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s45, v16
; VI-NEXT:    s_lshr_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v9, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s44, v16
; VI-NEXT:    s_lshr_b32 s4, s43, 16
; VI-NEXT:    v_or_b32_e32 v8, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s43, v16
; VI-NEXT:    s_lshr_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v7, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s42, v16
; VI-NEXT:    s_lshr_b32 s4, s41, 16
; VI-NEXT:    v_or_b32_e32 v6, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s41, v16
; VI-NEXT:    s_lshr_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v5, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s40, v16
; VI-NEXT:    s_lshr_b32 s4, s39, 16
; VI-NEXT:    v_or_b32_e32 v4, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v2, v2, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v3, s50, v16
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s39, v16
; VI-NEXT:    s_lshr_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v14, v3, v2
; VI-NEXT:    v_or_b32_e32 v3, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s38, v16
; VI-NEXT:    s_lshr_b32 s4, s37, 16
; VI-NEXT:    v_or_b32_e32 v2, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v1, s37, v16
; VI-NEXT:    s_lshr_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v1, v1, v0
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    v_add_f16_sdwa v0, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v17, s36, v16
; VI-NEXT:    s_lshr_b32 s4, s67, 16
; VI-NEXT:    v_or_b32_e32 v0, v17, v0
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s67, v16
; VI-NEXT:    s_lshr_b32 s4, s66, 16
; VI-NEXT:    v_or_b32_e32 v31, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s66, v16
; VI-NEXT:    s_lshr_b32 s4, s65, 16
; VI-NEXT:    v_or_b32_e32 v30, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s65, v16
; VI-NEXT:    s_lshr_b32 s4, s64, 16
; VI-NEXT:    v_or_b32_e32 v29, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s64, v16
; VI-NEXT:    s_lshr_b32 s4, s63, 16
; VI-NEXT:    v_or_b32_e32 v28, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s63, v16
; VI-NEXT:    s_lshr_b32 s4, s62, 16
; VI-NEXT:    v_or_b32_e32 v27, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s62, v16
; VI-NEXT:    s_lshr_b32 s4, s61, 16
; VI-NEXT:    v_or_b32_e32 v26, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s61, v16
; VI-NEXT:    s_lshr_b32 s4, s60, 16
; VI-NEXT:    v_or_b32_e32 v25, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s60, v16
; VI-NEXT:    s_lshr_b32 s4, s59, 16
; VI-NEXT:    v_or_b32_e32 v24, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s59, v16
; VI-NEXT:    s_lshr_b32 s4, s58, 16
; VI-NEXT:    v_or_b32_e32 v23, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s58, v16
; VI-NEXT:    s_lshr_b32 s4, s57, 16
; VI-NEXT:    v_or_b32_e32 v22, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s57, v16
; VI-NEXT:    s_lshr_b32 s4, s56, 16
; VI-NEXT:    v_or_b32_e32 v21, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s56, v16
; VI-NEXT:    s_lshr_b32 s4, s55, 16
; VI-NEXT:    v_or_b32_e32 v20, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s55, v16
; VI-NEXT:    s_lshr_b32 s4, s54, 16
; VI-NEXT:    v_or_b32_e32 v19, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s54, v16
; VI-NEXT:    s_lshr_b32 s4, s53, 16
; VI-NEXT:    v_or_b32_e32 v18, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v32, s53, v16
; VI-NEXT:    s_lshr_b32 s4, s52, 16
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_mov_b32_e32 v32, s4
; VI-NEXT:    v_add_f16_sdwa v32, v32, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, s52, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:    s_branch .LBB83_5
; VI-NEXT:  .LBB83_3:
; VI-NEXT:    s_branch .LBB83_2
; VI-NEXT:  .LBB83_4:
; VI-NEXT:    v_mov_b32_e32 v0, s36
; VI-NEXT:    v_mov_b32_e32 v1, s37
; VI-NEXT:    v_mov_b32_e32 v2, s38
; VI-NEXT:    v_mov_b32_e32 v3, s39
; VI-NEXT:    v_mov_b32_e32 v4, s40
; VI-NEXT:    v_mov_b32_e32 v5, s41
; VI-NEXT:    v_mov_b32_e32 v6, s42
; VI-NEXT:    v_mov_b32_e32 v7, s43
; VI-NEXT:    v_mov_b32_e32 v8, s44
; VI-NEXT:    v_mov_b32_e32 v9, s45
; VI-NEXT:    v_mov_b32_e32 v10, s46
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_mov_b32_e32 v12, s48
; VI-NEXT:    v_mov_b32_e32 v13, s49
; VI-NEXT:    v_mov_b32_e32 v14, s50
; VI-NEXT:    v_mov_b32_e32 v15, s51
; VI-NEXT:    v_mov_b32_e32 v16, s52
; VI-NEXT:    v_mov_b32_e32 v17, s53
; VI-NEXT:    v_mov_b32_e32 v18, s54
; VI-NEXT:    v_mov_b32_e32 v19, s55
; VI-NEXT:    v_mov_b32_e32 v20, s56
; VI-NEXT:    v_mov_b32_e32 v21, s57
; VI-NEXT:    v_mov_b32_e32 v22, s58
; VI-NEXT:    v_mov_b32_e32 v23, s59
; VI-NEXT:    v_mov_b32_e32 v24, s60
; VI-NEXT:    v_mov_b32_e32 v25, s61
; VI-NEXT:    v_mov_b32_e32 v26, s62
; VI-NEXT:    v_mov_b32_e32 v27, s63
; VI-NEXT:    v_mov_b32_e32 v28, s64
; VI-NEXT:    v_mov_b32_e32 v29, s65
; VI-NEXT:    v_mov_b32_e32 v30, s66
; VI-NEXT:    v_mov_b32_e32 v31, s67
; VI-NEXT:  .LBB83_5: ; %end
; VI-NEXT:    v_readlane_b32 s67, v33, 15
; VI-NEXT:    v_readlane_b32 s66, v33, 14
; VI-NEXT:    v_readlane_b32 s65, v33, 13
; VI-NEXT:    v_readlane_b32 s64, v33, 12
; VI-NEXT:    v_readlane_b32 s55, v33, 11
; VI-NEXT:    v_readlane_b32 s54, v33, 10
; VI-NEXT:    v_readlane_b32 s53, v33, 9
; VI-NEXT:    v_readlane_b32 s52, v33, 8
; VI-NEXT:    v_readlane_b32 s51, v33, 7
; VI-NEXT:    v_readlane_b32 s50, v33, 6
; VI-NEXT:    v_readlane_b32 s49, v33, 5
; VI-NEXT:    v_readlane_b32 s48, v33, 4
; VI-NEXT:    v_readlane_b32 s39, v33, 3
; VI-NEXT:    v_readlane_b32 s38, v33, 2
; VI-NEXT:    v_readlane_b32 s37, v33, 1
; VI-NEXT:    v_readlane_b32 s36, v33, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v16f64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB83_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB83_4
; GFX9-NEXT:  .LBB83_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, s51, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, s50, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, s49, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, s48, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, s47, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, s46, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, s45, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, s44, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, s43, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, s42, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, s41, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, s40, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, s39, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, s38, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, s37, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, s36, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v31, s67, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, s66, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, s65, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, s64, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, s63, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, s62, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, s61, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, s60, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, s59, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, s58, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, s57, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, s56, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, s55, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, s54, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, s53, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, s52, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB83_5
; GFX9-NEXT:  .LBB83_3:
; GFX9-NEXT:    s_branch .LBB83_2
; GFX9-NEXT:  .LBB83_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB83_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v16f64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB83_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB83_4
; GFX11-NEXT:  .LBB83_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s51 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s50 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s49 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s48 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s47 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s46 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s45 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s44 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s43 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s42 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s41 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s40 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s39 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s38 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s37 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s36 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, s67 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, s66 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s65 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s64 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, s63 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, s62 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, s61 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s60 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s59 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, s58 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s57 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s56 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s55 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s54 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s53 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s52 op_sel_hi:[0,1]
; GFX11-NEXT:    s_branch .LBB83_5
; GFX11-NEXT:  .LBB83_3:
; GFX11-NEXT:    s_branch .LBB83_2
; GFX11-NEXT:  .LBB83_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB83_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define <64 x i16> @bitcast_v16f64_to_v64i16(<16 x double> %a, i32 %b) {
; SI-LABEL: bitcast_v16f64_to_v64i16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB84_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v40, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v43, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v46, v3, v2, 16
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_alignbit_b32 v57, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB84_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB84_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; SI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; SI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; SI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; SI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; SI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; SI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; SI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; SI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; SI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; SI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; SI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; SI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; SI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; SI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; SI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; SI-NEXT:    v_alignbit_b32 v32, v31, v30, 16
; SI-NEXT:    v_alignbit_b32 v33, v29, v28, 16
; SI-NEXT:    v_alignbit_b32 v34, v27, v26, 16
; SI-NEXT:    v_alignbit_b32 v35, v25, v24, 16
; SI-NEXT:    v_alignbit_b32 v36, v23, v22, 16
; SI-NEXT:    v_alignbit_b32 v37, v21, v20, 16
; SI-NEXT:    v_alignbit_b32 v38, v19, v18, 16
; SI-NEXT:    v_alignbit_b32 v39, v17, v16, 16
; SI-NEXT:    v_alignbit_b32 v48, v15, v14, 16
; SI-NEXT:    v_alignbit_b32 v49, v13, v12, 16
; SI-NEXT:    v_alignbit_b32 v50, v11, v10, 16
; SI-NEXT:    v_alignbit_b32 v52, v9, v8, 16
; SI-NEXT:    v_alignbit_b32 v40, v7, v6, 16
; SI-NEXT:    v_alignbit_b32 v43, v5, v4, 16
; SI-NEXT:    v_alignbit_b32 v46, v3, v2, 16
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_alignbit_b32 v57, v1, v0, 16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v13
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v9
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v7
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v5
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; SI-NEXT:  .LBB84_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v57
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v46
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v43
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; SI-NEXT:    v_or_b32_e32 v0, v0, v57
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v63
; SI-NEXT:    v_or_b32_e32 v2, v2, v46
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v43
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v61
; SI-NEXT:    v_or_b32_e32 v6, v6, v40
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v60
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_or_b32_e32 v1, v1, v57
; SI-NEXT:    v_or_b32_e32 v3, v3, v46
; SI-NEXT:    v_or_b32_e32 v5, v5, v43
; SI-NEXT:    v_or_b32_e32 v7, v7, v40
; SI-NEXT:    v_or_b32_e32 v8, v8, v52
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v49
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v56
; SI-NEXT:    v_or_b32_e32 v14, v14, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v47
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v45
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v44
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v42
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v41
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v55
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v54
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v53
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v9, v9, v52
; SI-NEXT:    v_or_b32_e32 v11, v11, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v49
; SI-NEXT:    v_or_b32_e32 v15, v15, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v64i16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB84_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; VI-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; VI-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; VI-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; VI-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; VI-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; VI-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; VI-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; VI-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; VI-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; VI-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; VI-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; VI-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; VI-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; VI-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; VI-NEXT:  .LBB84_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v64i16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB84_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX9-NEXT:  .LBB84_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v64i16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB84_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[14:15], v[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], v[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], v[10:11], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], v[8:9], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], v[6:7], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], v[4:5], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], v[2:3], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], v[0:1], 1.0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_add_f64 v[30:31], v[30:31], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], v[28:29], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], v[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], v[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], v[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], v[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], v[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], v[16:17], 1.0
; GFX11-NEXT:  .LBB84_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define inreg <64 x i16> @bitcast_v16f64_to_v64i16_scalar(<16 x double> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v16f64_to_v64i16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_readfirstlane_b32 s45, v17
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    v_readfirstlane_b32 s43, v15
; SI-NEXT:    v_readfirstlane_b32 s42, v14
; SI-NEXT:    v_readfirstlane_b32 s41, v13
; SI-NEXT:    v_readfirstlane_b32 s40, v12
; SI-NEXT:    v_readfirstlane_b32 s15, v11
; SI-NEXT:    v_readfirstlane_b32 s14, v10
; SI-NEXT:    v_readfirstlane_b32 s13, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s11, v7
; SI-NEXT:    v_readfirstlane_b32 s10, v6
; SI-NEXT:    v_readfirstlane_b32 s9, v5
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_readfirstlane_b32 s7, v3
; SI-NEXT:    v_readfirstlane_b32 s6, v2
; SI-NEXT:    v_readfirstlane_b32 s5, v1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_readfirstlane_b32 s4, v0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB85_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshr_b32 s69, s45, 16
; SI-NEXT:    s_lshr_b32 s68, s43, 16
; SI-NEXT:    s_lshr_b32 s67, s41, 16
; SI-NEXT:    s_lshr_b32 s66, s15, 16
; SI-NEXT:    s_lshr_b32 s65, s13, 16
; SI-NEXT:    s_lshr_b32 s64, s11, 16
; SI-NEXT:    s_lshr_b32 s55, s9, 16
; SI-NEXT:    s_lshr_b32 s54, s7, 16
; SI-NEXT:    s_lshr_b32 s53, s5, 16
; SI-NEXT:    s_lshr_b32 s52, s29, 16
; SI-NEXT:    s_lshr_b32 s51, s27, 16
; SI-NEXT:    s_lshr_b32 s50, s25, 16
; SI-NEXT:    s_lshr_b32 s49, s23, 16
; SI-NEXT:    s_lshr_b32 s48, s21, 16
; SI-NEXT:    s_lshr_b32 s39, s19, 16
; SI-NEXT:    s_lshr_b32 s38, s17, 16
; SI-NEXT:    s_lshr_b64 s[46:47], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[60:61], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[56:57], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[58:59], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[62:63], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[72:73], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[74:75], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[4:5], 16
; SI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 16
; SI-NEXT:    s_lshr_b64 s[92:93], s[24:25], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[22:23], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[20:21], 16
; SI-NEXT:    s_lshr_b64 s[34:35], s[18:19], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[16:17], 16
; SI-NEXT:    s_cbranch_execnz .LBB85_4
; SI-NEXT:  .LBB85_2: ; %cmp.true
; SI-NEXT:    v_add_f64 v[28:29], s[42:43], 1.0
; SI-NEXT:    v_add_f64 v[26:27], s[40:41], 1.0
; SI-NEXT:    v_add_f64 v[30:31], s[44:45], 1.0
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v29
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v27
; SI-NEXT:    v_add_f64 v[16:17], s[6:7], 1.0
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[32:33], v[30:31], 16
; SI-NEXT:    v_add_f64 v[24:25], s[14:15], 1.0
; SI-NEXT:    v_add_f64 v[14:15], s[4:5], 1.0
; SI-NEXT:    v_lshr_b64 v[33:34], v[28:29], 16
; SI-NEXT:    v_add_f64 v[22:23], s[12:13], 1.0
; SI-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; SI-NEXT:    v_lshr_b64 v[34:35], v[26:27], 16
; SI-NEXT:    v_lshr_b64 v[48:49], v[16:17], 16
; SI-NEXT:    v_add_f64 v[20:21], s[10:11], 1.0
; SI-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; SI-NEXT:    v_lshr_b64 v[35:36], v[24:25], 16
; SI-NEXT:    v_lshr_b64 v[49:50], v[14:15], 16
; SI-NEXT:    v_add_f64 v[18:19], s[8:9], 1.0
; SI-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; SI-NEXT:    v_lshr_b64 v[36:37], v[22:23], 16
; SI-NEXT:    v_lshr_b64 v[50:51], v[12:13], 16
; SI-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; SI-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; SI-NEXT:    v_lshr_b64 v[37:38], v[20:21], 16
; SI-NEXT:    v_lshr_b64 v[51:52], v[10:11], 16
; SI-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; SI-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; SI-NEXT:    v_lshr_b64 v[38:39], v[18:19], 16
; SI-NEXT:    v_lshr_b64 v[52:53], v[8:9], 16
; SI-NEXT:    v_lshr_b64 v[53:54], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[39:40], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[54:55], v[4:5], 16
; SI-NEXT:    v_lshr_b64 v[40:41], v[0:1], 16
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v1
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v31
; SI-NEXT:    s_branch .LBB85_5
; SI-NEXT:  .LBB85_3:
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    s_branch .LBB85_2
; SI-NEXT:  .LBB85_4:
; SI-NEXT:    v_mov_b32_e32 v32, s68
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v32, s67
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v31, s45
; SI-NEXT:    v_mov_b32_e32 v29, s43
; SI-NEXT:    v_mov_b32_e32 v27, s41
; SI-NEXT:    v_mov_b32_e32 v25, s15
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v21, s11
; SI-NEXT:    v_mov_b32_e32 v19, s9
; SI-NEXT:    v_mov_b32_e32 v17, s7
; SI-NEXT:    v_mov_b32_e32 v15, s5
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v14, s4
; SI-NEXT:    v_mov_b32_e32 v16, s6
; SI-NEXT:    v_mov_b32_e32 v18, s8
; SI-NEXT:    v_mov_b32_e32 v20, s10
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v24, s14
; SI-NEXT:    v_mov_b32_e32 v26, s40
; SI-NEXT:    v_mov_b32_e32 v28, s42
; SI-NEXT:    v_mov_b32_e32 v30, s44
; SI-NEXT:    v_mov_b32_e32 v41, s69
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v45, s66
; SI-NEXT:    v_mov_b32_e32 v46, s65
; SI-NEXT:    v_mov_b32_e32 v47, s64
; SI-NEXT:    v_mov_b32_e32 v56, s55
; SI-NEXT:    v_mov_b32_e32 v57, s54
; SI-NEXT:    v_mov_b32_e32 v58, s53
; SI-NEXT:    v_mov_b32_e32 v59, s52
; SI-NEXT:    v_mov_b32_e32 v60, s51
; SI-NEXT:    v_mov_b32_e32 v61, s50
; SI-NEXT:    v_mov_b32_e32 v62, s49
; SI-NEXT:    v_mov_b32_e32 v42, s48
; SI-NEXT:    v_mov_b32_e32 v43, s39
; SI-NEXT:    v_mov_b32_e32 v44, s38
; SI-NEXT:    v_mov_b32_e32 v40, s36
; SI-NEXT:    v_mov_b32_e32 v39, s34
; SI-NEXT:    v_mov_b32_e32 v54, s30
; SI-NEXT:    v_mov_b32_e32 v53, s94
; SI-NEXT:    v_mov_b32_e32 v52, s92
; SI-NEXT:    v_mov_b32_e32 v51, s90
; SI-NEXT:    v_mov_b32_e32 v50, s88
; SI-NEXT:    v_mov_b32_e32 v49, s78
; SI-NEXT:    v_mov_b32_e32 v48, s74
; SI-NEXT:    v_mov_b32_e32 v38, s72
; SI-NEXT:    v_mov_b32_e32 v37, s62
; SI-NEXT:    v_mov_b32_e32 v36, s58
; SI-NEXT:    v_mov_b32_e32 v35, s56
; SI-NEXT:    v_mov_b32_e32 v34, s76
; SI-NEXT:    v_mov_b32_e32 v33, s60
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v32, s46
; SI-NEXT:  .LBB85_5: ; %end
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v39
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v43
; SI-NEXT:    v_or_b32_e32 v3, v3, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v54
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v39
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v42
; SI-NEXT:    v_or_b32_e32 v5, v5, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v53
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v39
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v62
; SI-NEXT:    v_or_b32_e32 v7, v7, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v52
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v39
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v61
; SI-NEXT:    v_or_b32_e32 v9, v9, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v51
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v39
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v60
; SI-NEXT:    v_or_b32_e32 v11, v11, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v50
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v39
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v59
; SI-NEXT:    v_or_b32_e32 v13, v13, v39
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v49
; SI-NEXT:    v_or_b32_e32 v14, v14, v39
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v58
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v40
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v15, v15, v39
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v48
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v0, v0, v55
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v44
; SI-NEXT:    v_or_b32_e32 v16, v16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v57
; SI-NEXT:    v_or_b32_e32 v18, v18, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v56
; SI-NEXT:    v_or_b32_e32 v20, v20, v37
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v47
; SI-NEXT:    v_or_b32_e32 v22, v22, v36
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v46
; SI-NEXT:    v_or_b32_e32 v24, v24, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v45
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v41
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_or_b32_e32 v26, v26, v34
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v28, v28, v33
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v1, v1, v55
; SI-NEXT:    v_or_b32_e32 v17, v17, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v38
; SI-NEXT:    v_or_b32_e32 v21, v21, v37
; SI-NEXT:    v_or_b32_e32 v23, v23, v36
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_or_b32_e32 v27, v27, v34
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_or_b32_e32 v29, v29, v33
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v16f64_to_v64i16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s36, 2
; VI-NEXT:    v_writelane_b32 v32, s37, 3
; VI-NEXT:    v_writelane_b32 v32, s38, 4
; VI-NEXT:    v_writelane_b32 v32, s39, 5
; VI-NEXT:    v_writelane_b32 v32, s48, 6
; VI-NEXT:    v_writelane_b32 v32, s49, 7
; VI-NEXT:    v_writelane_b32 v32, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB85_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB85_4
; VI-NEXT:  .LBB85_2: ; %cmp.true
; VI-NEXT:    v_add_f64 v[14:15], s[30:31], 1.0
; VI-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; VI-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; VI-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; VI-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; VI-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; VI-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; VI-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; VI-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; VI-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; VI-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; VI-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; VI-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; VI-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; VI-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; VI-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; VI-NEXT:    s_branch .LBB85_5
; VI-NEXT:  .LBB85_3:
; VI-NEXT:    s_branch .LBB85_2
; VI-NEXT:  .LBB85_4:
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:  .LBB85_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v32, 9
; VI-NEXT:    v_readlane_b32 s50, v32, 8
; VI-NEXT:    v_readlane_b32 s49, v32, 7
; VI-NEXT:    v_readlane_b32 s48, v32, 6
; VI-NEXT:    v_readlane_b32 s39, v32, 5
; VI-NEXT:    v_readlane_b32 s38, v32, 4
; VI-NEXT:    v_readlane_b32 s37, v32, 3
; VI-NEXT:    v_readlane_b32 s36, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v16f64_to_v64i16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB85_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB85_4
; GFX9-NEXT:  .LBB85_2: ; %cmp.true
; GFX9-NEXT:    v_add_f64 v[14:15], s[30:31], 1.0
; GFX9-NEXT:    v_add_f64 v[12:13], s[28:29], 1.0
; GFX9-NEXT:    v_add_f64 v[10:11], s[26:27], 1.0
; GFX9-NEXT:    v_add_f64 v[8:9], s[24:25], 1.0
; GFX9-NEXT:    v_add_f64 v[6:7], s[22:23], 1.0
; GFX9-NEXT:    v_add_f64 v[4:5], s[20:21], 1.0
; GFX9-NEXT:    v_add_f64 v[2:3], s[18:19], 1.0
; GFX9-NEXT:    v_add_f64 v[0:1], s[16:17], 1.0
; GFX9-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; GFX9-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; GFX9-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; GFX9-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; GFX9-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; GFX9-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; GFX9-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; GFX9-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; GFX9-NEXT:    s_branch .LBB85_5
; GFX9-NEXT:  .LBB85_3:
; GFX9-NEXT:    s_branch .LBB85_2
; GFX9-NEXT:  .LBB85_4:
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:  .LBB85_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v16f64_to_v64i16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB85_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB85_4
; GFX11-NEXT:  .LBB85_2: ; %cmp.true
; GFX11-NEXT:    v_add_f64 v[14:15], s[26:27], 1.0
; GFX11-NEXT:    v_add_f64 v[12:13], s[24:25], 1.0
; GFX11-NEXT:    v_add_f64 v[10:11], s[22:23], 1.0
; GFX11-NEXT:    v_add_f64 v[8:9], s[20:21], 1.0
; GFX11-NEXT:    v_add_f64 v[6:7], s[18:19], 1.0
; GFX11-NEXT:    v_add_f64 v[4:5], s[16:17], 1.0
; GFX11-NEXT:    v_add_f64 v[2:3], s[14:15], 1.0
; GFX11-NEXT:    v_add_f64 v[0:1], s[12:13], 1.0
; GFX11-NEXT:    v_add_f64 v[30:31], s[50:51], 1.0
; GFX11-NEXT:    v_add_f64 v[28:29], s[48:49], 1.0
; GFX11-NEXT:    v_add_f64 v[26:27], s[46:47], 1.0
; GFX11-NEXT:    v_add_f64 v[24:25], s[44:45], 1.0
; GFX11-NEXT:    v_add_f64 v[22:23], s[42:43], 1.0
; GFX11-NEXT:    v_add_f64 v[20:21], s[40:41], 1.0
; GFX11-NEXT:    v_add_f64 v[18:19], s[38:39], 1.0
; GFX11-NEXT:    v_add_f64 v[16:17], s[36:37], 1.0
; GFX11-NEXT:    s_branch .LBB85_5
; GFX11-NEXT:  .LBB85_3:
; GFX11-NEXT:    s_branch .LBB85_2
; GFX11-NEXT:  .LBB85_4:
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:  .LBB85_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <16 x double> %a, splat (double 1.000000e+00)
  %a2 = bitcast <16 x double> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <16 x double> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define <16 x double> @bitcast_v64i16_to_v16f64(<64 x i16> %a, i32 %b) {
; SI-LABEL: bitcast_v64i16_to_v16f64:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v62, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v45, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v19
; SI-NEXT:    v_mov_b32_e32 v46, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v20
; SI-NEXT:    v_mov_b32_e32 v47, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v21
; SI-NEXT:    v_mov_b32_e32 v56, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v22
; SI-NEXT:    v_mov_b32_e32 v57, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v23
; SI-NEXT:    v_mov_b32_e32 v58, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v24
; SI-NEXT:    v_mov_b32_e32 v59, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v25
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    v_mov_b32_e32 v61, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v27
; SI-NEXT:    v_mov_b32_e32 v43, v11
; SI-NEXT:    v_mov_b32_e32 v44, v10
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v18
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v17
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v15
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v43
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v45
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v46
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v59
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v61
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v27
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v26
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v25
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v23
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v22
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v21
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v20
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v19
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v18
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v17
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v31
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v10
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v62
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v28
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v29
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v30
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v63
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB86_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v46
; SI-NEXT:    v_or_b32_e32 v8, v8, v32
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v62
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v60
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v59
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v58
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v57
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v56
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v47
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v45
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v44
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    v_or_b32_e32 v0, v0, v49
; SI-NEXT:    v_or_b32_e32 v1, v1, v42
; SI-NEXT:    v_or_b32_e32 v2, v2, v34
; SI-NEXT:    v_or_b32_e32 v3, v3, v41
; SI-NEXT:    v_or_b32_e32 v4, v4, v48
; SI-NEXT:    v_or_b32_e32 v5, v5, v40
; SI-NEXT:    v_or_b32_e32 v6, v6, v33
; SI-NEXT:    v_or_b32_e32 v7, v7, v55
; SI-NEXT:    v_or_b32_e32 v9, v9, v54
; SI-NEXT:    v_or_b32_e32 v10, v10, v39
; SI-NEXT:    v_or_b32_e32 v11, v11, v53
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; kill: killed $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v13, v52
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v37
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v15, v51
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v36
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v17, v50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v18, v35
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v63
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:  .LBB86_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB86_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v46
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v32, v8
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v62
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v61
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v60
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v59
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v58
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v57
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v56
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v47
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v45
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v44
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_or_b32_e32 v1, v42, v1
; SI-NEXT:    s_mov_b32 s6, 0x30000
; SI-NEXT:    v_or_b32_e32 v2, v34, v2
; SI-NEXT:    v_or_b32_e32 v3, v41, v3
; SI-NEXT:    v_or_b32_e32 v4, v48, v4
; SI-NEXT:    v_or_b32_e32 v5, v40, v5
; SI-NEXT:    v_or_b32_e32 v6, v33, v6
; SI-NEXT:    v_or_b32_e32 v7, v55, v7
; SI-NEXT:    v_or_b32_e32 v9, v54, v9
; SI-NEXT:    v_or_b32_e32 v10, v39, v10
; SI-NEXT:    v_or_b32_e32 v11, v53, v11
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x30000, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v23, v22
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v38, v12
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v25, v24
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v52, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v37, v14
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v27, v26
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v15, v51, v15
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v36, v16
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v17, v50, v17
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_add_i32_e32 v28, vcc, s6, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v35, v18
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_add_i32_e32 v29, vcc, 0x30000, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v63
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v30, vcc, 0x30000, v30
; SI-NEXT:    v_add_i32_e32 v31, vcc, 0x30000, v31
; SI-NEXT:  .LBB86_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v16f64:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB86_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v33, 3
; VI-NEXT:    v_add_u16_e32 v32, 3, v15
; VI-NEXT:    v_add_u16_sdwa v15, v15, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v32, v15
; VI-NEXT:    v_add_u16_e32 v32, 3, v14
; VI-NEXT:    v_add_u16_sdwa v14, v14, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v32, v14
; VI-NEXT:    v_add_u16_e32 v32, 3, v13
; VI-NEXT:    v_add_u16_sdwa v13, v13, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v32, v13
; VI-NEXT:    v_add_u16_e32 v32, 3, v12
; VI-NEXT:    v_add_u16_sdwa v12, v12, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v32, v12
; VI-NEXT:    v_add_u16_e32 v32, 3, v11
; VI-NEXT:    v_add_u16_sdwa v11, v11, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v32, v11
; VI-NEXT:    v_add_u16_e32 v32, 3, v10
; VI-NEXT:    v_add_u16_sdwa v10, v10, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v32, v10
; VI-NEXT:    v_add_u16_e32 v32, 3, v9
; VI-NEXT:    v_add_u16_sdwa v9, v9, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v32, v9
; VI-NEXT:    v_add_u16_e32 v32, 3, v8
; VI-NEXT:    v_add_u16_sdwa v8, v8, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v32, v8
; VI-NEXT:    v_add_u16_e32 v32, 3, v7
; VI-NEXT:    v_add_u16_sdwa v7, v7, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v32, v7
; VI-NEXT:    v_add_u16_e32 v32, 3, v6
; VI-NEXT:    v_add_u16_sdwa v6, v6, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v32, v6
; VI-NEXT:    v_add_u16_e32 v32, 3, v5
; VI-NEXT:    v_add_u16_sdwa v5, v5, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v32, v5
; VI-NEXT:    v_add_u16_e32 v32, 3, v4
; VI-NEXT:    v_add_u16_sdwa v4, v4, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v32, v4
; VI-NEXT:    v_add_u16_e32 v32, 3, v3
; VI-NEXT:    v_add_u16_sdwa v3, v3, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v32, v3
; VI-NEXT:    v_add_u16_e32 v32, 3, v2
; VI-NEXT:    v_add_u16_sdwa v2, v2, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v32, v2
; VI-NEXT:    v_add_u16_e32 v32, 3, v1
; VI-NEXT:    v_add_u16_sdwa v1, v1, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v32, v1
; VI-NEXT:    v_add_u16_e32 v32, 3, v0
; VI-NEXT:    v_add_u16_sdwa v0, v0, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v32, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v31
; VI-NEXT:    v_add_u16_sdwa v31, v31, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v32, v31
; VI-NEXT:    v_add_u16_e32 v32, 3, v30
; VI-NEXT:    v_add_u16_sdwa v30, v30, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v32, v30
; VI-NEXT:    v_add_u16_e32 v32, 3, v29
; VI-NEXT:    v_add_u16_sdwa v29, v29, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v32, v29
; VI-NEXT:    v_add_u16_e32 v32, 3, v28
; VI-NEXT:    v_add_u16_sdwa v28, v28, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v32, v28
; VI-NEXT:    v_add_u16_e32 v32, 3, v27
; VI-NEXT:    v_add_u16_sdwa v27, v27, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v32, v27
; VI-NEXT:    v_add_u16_e32 v32, 3, v26
; VI-NEXT:    v_add_u16_sdwa v26, v26, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v32, v26
; VI-NEXT:    v_add_u16_e32 v32, 3, v25
; VI-NEXT:    v_add_u16_sdwa v25, v25, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v32, v25
; VI-NEXT:    v_add_u16_e32 v32, 3, v24
; VI-NEXT:    v_add_u16_sdwa v24, v24, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v32, v24
; VI-NEXT:    v_add_u16_e32 v32, 3, v23
; VI-NEXT:    v_add_u16_sdwa v23, v23, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v32, v23
; VI-NEXT:    v_add_u16_e32 v32, 3, v22
; VI-NEXT:    v_add_u16_sdwa v22, v22, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v32, v22
; VI-NEXT:    v_add_u16_e32 v32, 3, v21
; VI-NEXT:    v_add_u16_sdwa v21, v21, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v32, v21
; VI-NEXT:    v_add_u16_e32 v32, 3, v20
; VI-NEXT:    v_add_u16_sdwa v20, v20, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v32, v20
; VI-NEXT:    v_add_u16_e32 v32, 3, v19
; VI-NEXT:    v_add_u16_sdwa v19, v19, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v32, v19
; VI-NEXT:    v_add_u16_e32 v32, 3, v18
; VI-NEXT:    v_add_u16_sdwa v18, v18, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v32, v18
; VI-NEXT:    v_add_u16_e32 v32, 3, v17
; VI-NEXT:    v_add_u16_sdwa v17, v17, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_add_u16_sdwa v32, v16, v33 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB86_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64i16_to_v16f64:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB86_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB86_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v16f64:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB86_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:  .LBB86_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define inreg <16 x double> @bitcast_v64i16_to_v16f64_scalar(<64 x i16> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64i16_to_v16f64_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    v_writelane_b32 v32, s70, 22
; SI-NEXT:    v_writelane_b32 v32, s71, 23
; SI-NEXT:    v_writelane_b32 v32, s80, 24
; SI-NEXT:    v_writelane_b32 v32, s81, 25
; SI-NEXT:    v_writelane_b32 v32, s82, 26
; SI-NEXT:    v_writelane_b32 v32, s83, 27
; SI-NEXT:    v_writelane_b32 v32, s84, 28
; SI-NEXT:    v_writelane_b32 v32, s85, 29
; SI-NEXT:    v_writelane_b32 v32, s86, 30
; SI-NEXT:    v_writelane_b32 v32, s87, 31
; SI-NEXT:    v_writelane_b32 v32, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s9, v16
; SI-NEXT:    v_writelane_b32 v32, s97, 33
; SI-NEXT:    s_lshr_b32 s14, s9, 16
; SI-NEXT:    v_readfirstlane_b32 s13, v14
; SI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v32, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s7, v17
; SI-NEXT:    v_readfirstlane_b32 s11, v15
; SI-NEXT:    s_lshr_b32 s72, s13, 16
; SI-NEXT:    v_readfirstlane_b32 s15, v13
; SI-NEXT:    v_readfirstlane_b32 s73, v12
; SI-NEXT:    v_readfirstlane_b32 s75, v11
; SI-NEXT:    v_readfirstlane_b32 s77, v10
; SI-NEXT:    v_readfirstlane_b32 s79, v9
; SI-NEXT:    v_readfirstlane_b32 s89, v8
; SI-NEXT:    v_readfirstlane_b32 s91, v7
; SI-NEXT:    v_readfirstlane_b32 s94, v6
; SI-NEXT:    v_readfirstlane_b32 s31, v5
; SI-NEXT:    v_readfirstlane_b32 s68, v4
; SI-NEXT:    v_readfirstlane_b32 s80, v3
; SI-NEXT:    v_readfirstlane_b32 s83, v2
; SI-NEXT:    v_readfirstlane_b32 s86, v1
; SI-NEXT:    v_readfirstlane_b32 s97, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v33, s14, 0
; SI-NEXT:    v_writelane_b32 v32, s99, 35
; SI-NEXT:    s_lshr_b32 s92, s29, 16
; SI-NEXT:    s_lshr_b32 s95, s28, 16
; SI-NEXT:    s_lshr_b32 s34, s27, 16
; SI-NEXT:    s_lshr_b32 s69, s26, 16
; SI-NEXT:    s_lshr_b32 s71, s25, 16
; SI-NEXT:    s_lshr_b32 s81, s24, 16
; SI-NEXT:    s_lshr_b32 s84, s23, 16
; SI-NEXT:    s_lshr_b32 s87, s22, 16
; SI-NEXT:    s_lshr_b32 s98, s21, 16
; SI-NEXT:    s_lshr_b32 s99, s20, 16
; SI-NEXT:    s_lshr_b32 s6, s19, 16
; SI-NEXT:    s_lshr_b32 s8, s18, 16
; SI-NEXT:    s_lshr_b32 s10, s17, 16
; SI-NEXT:    s_lshr_b32 s12, s16, 16
; SI-NEXT:    s_lshr_b32 vcc_lo, s7, 16
; SI-NEXT:    s_lshr_b32 vcc_hi, s11, 16
; SI-NEXT:    s_lshr_b32 s63, s15, 16
; SI-NEXT:    s_lshr_b32 s62, s73, 16
; SI-NEXT:    s_lshr_b32 s61, s75, 16
; SI-NEXT:    s_lshr_b32 s60, s77, 16
; SI-NEXT:    s_lshr_b32 s78, s79, 16
; SI-NEXT:    s_lshr_b32 s88, s89, 16
; SI-NEXT:    s_lshr_b32 s90, s91, 16
; SI-NEXT:    s_lshr_b32 s93, s94, 16
; SI-NEXT:    s_lshr_b32 s30, s31, 16
; SI-NEXT:    s_lshr_b32 s35, s68, 16
; SI-NEXT:    s_lshr_b32 s70, s80, 16
; SI-NEXT:    s_lshr_b32 s82, s83, 16
; SI-NEXT:    s_lshr_b32 s85, s86, 16
; SI-NEXT:    s_lshr_b32 s96, s97, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v33, s72, 1
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v33, s63, 2
; SI-NEXT:    v_writelane_b32 v33, s62, 3
; SI-NEXT:    s_cbranch_scc0 .LBB87_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_or_b32 s36, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s38, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s10, 16
; SI-NEXT:    s_or_b32 s39, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s37, s40, s41
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s48, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s49, s4, s5
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s50, s4, s5
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s51, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s52, s4, s5
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s53, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s54, s4, s5
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s55, s4, s5
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s60, 16
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s61, 16
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s62, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s63, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s64, s4, s5
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s65, s4, s5
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s66, s4, s5
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    s_or_b32 s67, s4, s5
; SI-NEXT:    s_cbranch_execnz .LBB87_3
; SI-NEXT:  .LBB87_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_or_b32 s5, s10, s5
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s36, s4, 0x30000
; SI-NEXT:    s_add_i32 s37, s5, 0x30000
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s38, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s39, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s40, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s98, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s41, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s42, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s43, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s44, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s45, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s46, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s47, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s48, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s97, s97, 3
; SI-NEXT:    s_add_i32 s49, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s97, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s96, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s86, s86, 3
; SI-NEXT:    s_add_i32 s50, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s86, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s83, s83, 3
; SI-NEXT:    s_add_i32 s51, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s82, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s80, s80, 3
; SI-NEXT:    s_add_i32 s52, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s68, s68, 3
; SI-NEXT:    s_add_i32 s53, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s31, s31, 3
; SI-NEXT:    s_add_i32 s54, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s94, s94, 3
; SI-NEXT:    s_add_i32 s55, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s94, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s91, s91, 3
; SI-NEXT:    s_add_i32 s56, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s89, s89, 3
; SI-NEXT:    s_add_i32 s57, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s79, s79, 3
; SI-NEXT:    s_add_i32 s58, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s77, s77, 3
; SI-NEXT:    s_add_i32 s59, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s76, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s75, s75, 3
; SI-NEXT:    s_add_i32 s60, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s74, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s73, s73, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 3
; SI-NEXT:    s_add_i32 s61, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 2
; SI-NEXT:    s_add_i32 s62, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 1
; SI-NEXT:    s_add_i32 s63, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s64, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s72, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 0
; SI-NEXT:    s_add_i32 s65, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_add_i32 s66, s4, 0x30000
; SI-NEXT:    s_and_b32 s4, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_add_i32 s67, s4, 0x30000
; SI-NEXT:  .LBB87_3: ; %end
; SI-NEXT:    v_mov_b32_e32 v0, s36
; SI-NEXT:    v_mov_b32_e32 v1, s37
; SI-NEXT:    v_mov_b32_e32 v2, s38
; SI-NEXT:    v_mov_b32_e32 v3, s39
; SI-NEXT:    v_mov_b32_e32 v4, s40
; SI-NEXT:    v_mov_b32_e32 v5, s41
; SI-NEXT:    v_mov_b32_e32 v6, s42
; SI-NEXT:    v_mov_b32_e32 v7, s43
; SI-NEXT:    v_mov_b32_e32 v8, s44
; SI-NEXT:    v_mov_b32_e32 v9, s45
; SI-NEXT:    v_mov_b32_e32 v10, s46
; SI-NEXT:    v_mov_b32_e32 v11, s47
; SI-NEXT:    v_mov_b32_e32 v12, s48
; SI-NEXT:    v_mov_b32_e32 v13, s49
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_mov_b32_e32 v15, s51
; SI-NEXT:    v_mov_b32_e32 v16, s52
; SI-NEXT:    v_mov_b32_e32 v17, s53
; SI-NEXT:    v_mov_b32_e32 v18, s54
; SI-NEXT:    v_mov_b32_e32 v19, s55
; SI-NEXT:    v_mov_b32_e32 v20, s56
; SI-NEXT:    v_mov_b32_e32 v21, s57
; SI-NEXT:    v_mov_b32_e32 v22, s58
; SI-NEXT:    v_mov_b32_e32 v23, s59
; SI-NEXT:    v_mov_b32_e32 v24, s60
; SI-NEXT:    v_mov_b32_e32 v25, s61
; SI-NEXT:    v_mov_b32_e32 v26, s62
; SI-NEXT:    v_mov_b32_e32 v27, s63
; SI-NEXT:    v_mov_b32_e32 v28, s64
; SI-NEXT:    v_mov_b32_e32 v29, s65
; SI-NEXT:    v_mov_b32_e32 v30, s66
; SI-NEXT:    v_mov_b32_e32 v31, s67
; SI-NEXT:    v_readlane_b32 s99, v32, 35
; SI-NEXT:    v_readlane_b32 s98, v32, 34
; SI-NEXT:    v_readlane_b32 s97, v32, 33
; SI-NEXT:    v_readlane_b32 s96, v32, 32
; SI-NEXT:    v_readlane_b32 s87, v32, 31
; SI-NEXT:    v_readlane_b32 s86, v32, 30
; SI-NEXT:    v_readlane_b32 s85, v32, 29
; SI-NEXT:    v_readlane_b32 s84, v32, 28
; SI-NEXT:    v_readlane_b32 s83, v32, 27
; SI-NEXT:    v_readlane_b32 s82, v32, 26
; SI-NEXT:    v_readlane_b32 s81, v32, 25
; SI-NEXT:    v_readlane_b32 s80, v32, 24
; SI-NEXT:    v_readlane_b32 s71, v32, 23
; SI-NEXT:    v_readlane_b32 s70, v32, 22
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB87_4:
; SI-NEXT:    s_mov_b32 s76, s60
; SI-NEXT:    s_mov_b32 s74, s61
; SI-NEXT:    s_mov_b32 s72, vcc_hi
; SI-NEXT:    s_mov_b32 s14, vcc_lo
; SI-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
; SI-NEXT:    s_branch .LBB87_2
;
; VI-LABEL: bitcast_v64i16_to_v16f64_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    v_readfirstlane_b32 s46, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s47, v0
; VI-NEXT:    s_cbranch_scc0 .LBB87_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB87_3
; VI-NEXT:  .LBB87_2: ; %cmp.true
; VI-NEXT:    s_add_i32 s5, s46, 3
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s47, 3
; VI-NEXT:    s_add_i32 s46, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s29, 3
; VI-NEXT:    s_add_i32 s47, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s28, 3
; VI-NEXT:    s_add_i32 s29, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s28, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s27, 3
; VI-NEXT:    s_add_i32 s28, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s27, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s26, 3
; VI-NEXT:    s_add_i32 s27, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s26, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s25, 3
; VI-NEXT:    s_add_i32 s26, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s25, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s24, 3
; VI-NEXT:    s_add_i32 s25, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s24, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s23, 3
; VI-NEXT:    s_add_i32 s24, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s23, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s22, 3
; VI-NEXT:    s_add_i32 s23, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s22, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s21, 3
; VI-NEXT:    s_add_i32 s22, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s21, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s20, 3
; VI-NEXT:    s_add_i32 s21, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s20, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s19, 3
; VI-NEXT:    s_add_i32 s20, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s19, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s18, 3
; VI-NEXT:    s_add_i32 s19, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s18, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s17, 3
; VI-NEXT:    s_add_i32 s18, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s17, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s16, 3
; VI-NEXT:    s_add_i32 s17, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s16, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s6, 3
; VI-NEXT:    s_add_i32 s16, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s6, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s7, 3
; VI-NEXT:    s_add_i32 s6, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s7, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s8, 3
; VI-NEXT:    s_add_i32 s7, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s8, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s9, 3
; VI-NEXT:    s_add_i32 s8, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s9, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s10, 3
; VI-NEXT:    s_add_i32 s9, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s10, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s11, 3
; VI-NEXT:    s_add_i32 s10, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s11, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s12, 3
; VI-NEXT:    s_add_i32 s11, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s12, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s13, 3
; VI-NEXT:    s_add_i32 s12, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s13, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s14, 3
; VI-NEXT:    s_add_i32 s13, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s14, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s15, 3
; VI-NEXT:    s_add_i32 s14, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s15, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s40, 3
; VI-NEXT:    s_add_i32 s15, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s41, 3
; VI-NEXT:    s_add_i32 s40, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s42, 3
; VI-NEXT:    s_add_i32 s41, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s43, 3
; VI-NEXT:    s_add_i32 s42, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s44, 3
; VI-NEXT:    s_add_i32 s43, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s5, s45, 3
; VI-NEXT:    s_add_i32 s44, s4, 0x30000
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s45, s4, 0x30000
; VI-NEXT:  .LBB87_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s47
; VI-NEXT:    v_mov_b32_e32 v15, s46
; VI-NEXT:    v_mov_b32_e32 v16, s45
; VI-NEXT:    v_mov_b32_e32 v17, s44
; VI-NEXT:    v_mov_b32_e32 v18, s43
; VI-NEXT:    v_mov_b32_e32 v19, s42
; VI-NEXT:    v_mov_b32_e32 v20, s41
; VI-NEXT:    v_mov_b32_e32 v21, s40
; VI-NEXT:    v_mov_b32_e32 v22, s15
; VI-NEXT:    v_mov_b32_e32 v23, s14
; VI-NEXT:    v_mov_b32_e32 v24, s13
; VI-NEXT:    v_mov_b32_e32 v25, s12
; VI-NEXT:    v_mov_b32_e32 v26, s11
; VI-NEXT:    v_mov_b32_e32 v27, s10
; VI-NEXT:    v_mov_b32_e32 v28, s9
; VI-NEXT:    v_mov_b32_e32 v29, s8
; VI-NEXT:    v_mov_b32_e32 v30, s7
; VI-NEXT:    v_mov_b32_e32 v31, s6
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB87_4:
; VI-NEXT:    s_branch .LBB87_2
;
; GFX9-LABEL: bitcast_v64i16_to_v16f64_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s36, 0
; GFX9-NEXT:    v_writelane_b32 v32, s37, 1
; GFX9-NEXT:    v_writelane_b32 v32, s38, 2
; GFX9-NEXT:    v_writelane_b32 v32, s39, 3
; GFX9-NEXT:    v_writelane_b32 v32, s48, 4
; GFX9-NEXT:    v_writelane_b32 v32, s49, 5
; GFX9-NEXT:    v_writelane_b32 v32, s50, 6
; GFX9-NEXT:    v_writelane_b32 v32, s51, 7
; GFX9-NEXT:    v_writelane_b32 v32, s52, 8
; GFX9-NEXT:    v_writelane_b32 v32, s53, 9
; GFX9-NEXT:    v_writelane_b32 v32, s54, 10
; GFX9-NEXT:    v_writelane_b32 v32, s55, 11
; GFX9-NEXT:    v_writelane_b32 v32, s64, 12
; GFX9-NEXT:    v_writelane_b32 v32, s65, 13
; GFX9-NEXT:    v_writelane_b32 v32, s66, 14
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s67, 15
; GFX9-NEXT:    s_mov_b32 s49, s29
; GFX9-NEXT:    s_mov_b32 s48, s28
; GFX9-NEXT:    s_mov_b32 s47, s27
; GFX9-NEXT:    s_mov_b32 s46, s26
; GFX9-NEXT:    s_mov_b32 s45, s25
; GFX9-NEXT:    s_mov_b32 s44, s24
; GFX9-NEXT:    s_mov_b32 s43, s23
; GFX9-NEXT:    s_mov_b32 s42, s22
; GFX9-NEXT:    s_mov_b32 s41, s21
; GFX9-NEXT:    s_mov_b32 s40, s20
; GFX9-NEXT:    s_mov_b32 s39, s19
; GFX9-NEXT:    s_mov_b32 s38, s18
; GFX9-NEXT:    s_mov_b32 s37, s17
; GFX9-NEXT:    s_mov_b32 s36, s16
; GFX9-NEXT:    v_readfirstlane_b32 s67, v17
; GFX9-NEXT:    v_readfirstlane_b32 s66, v16
; GFX9-NEXT:    v_readfirstlane_b32 s65, v15
; GFX9-NEXT:    v_readfirstlane_b32 s64, v14
; GFX9-NEXT:    v_readfirstlane_b32 s63, v13
; GFX9-NEXT:    v_readfirstlane_b32 s62, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s60, v10
; GFX9-NEXT:    v_readfirstlane_b32 s59, v9
; GFX9-NEXT:    v_readfirstlane_b32 s58, v8
; GFX9-NEXT:    v_readfirstlane_b32 s57, v7
; GFX9-NEXT:    v_readfirstlane_b32 s56, v6
; GFX9-NEXT:    v_readfirstlane_b32 s55, v5
; GFX9-NEXT:    v_readfirstlane_b32 s54, v4
; GFX9-NEXT:    v_readfirstlane_b32 s53, v3
; GFX9-NEXT:    v_readfirstlane_b32 s52, v2
; GFX9-NEXT:    v_readfirstlane_b32 s51, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s50, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB87_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB87_4
; GFX9-NEXT:  .LBB87_2: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB87_5
; GFX9-NEXT:  .LBB87_3:
; GFX9-NEXT:    s_branch .LBB87_2
; GFX9-NEXT:  .LBB87_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s37
; GFX9-NEXT:    v_mov_b32_e32 v2, s38
; GFX9-NEXT:    v_mov_b32_e32 v3, s39
; GFX9-NEXT:    v_mov_b32_e32 v4, s40
; GFX9-NEXT:    v_mov_b32_e32 v5, s41
; GFX9-NEXT:    v_mov_b32_e32 v6, s42
; GFX9-NEXT:    v_mov_b32_e32 v7, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s44
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_mov_b32_e32 v10, s46
; GFX9-NEXT:    v_mov_b32_e32 v11, s47
; GFX9-NEXT:    v_mov_b32_e32 v12, s48
; GFX9-NEXT:    v_mov_b32_e32 v13, s49
; GFX9-NEXT:    v_mov_b32_e32 v14, s50
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    v_mov_b32_e32 v16, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s53
; GFX9-NEXT:    v_mov_b32_e32 v18, s54
; GFX9-NEXT:    v_mov_b32_e32 v19, s55
; GFX9-NEXT:    v_mov_b32_e32 v20, s56
; GFX9-NEXT:    v_mov_b32_e32 v21, s57
; GFX9-NEXT:    v_mov_b32_e32 v22, s58
; GFX9-NEXT:    v_mov_b32_e32 v23, s59
; GFX9-NEXT:    v_mov_b32_e32 v24, s60
; GFX9-NEXT:    v_mov_b32_e32 v25, s61
; GFX9-NEXT:    v_mov_b32_e32 v26, s62
; GFX9-NEXT:    v_mov_b32_e32 v27, s63
; GFX9-NEXT:    v_mov_b32_e32 v28, s64
; GFX9-NEXT:    v_mov_b32_e32 v29, s65
; GFX9-NEXT:    v_mov_b32_e32 v30, s66
; GFX9-NEXT:    v_mov_b32_e32 v31, s67
; GFX9-NEXT:  .LBB87_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s67, v32, 15
; GFX9-NEXT:    v_readlane_b32 s66, v32, 14
; GFX9-NEXT:    v_readlane_b32 s65, v32, 13
; GFX9-NEXT:    v_readlane_b32 s64, v32, 12
; GFX9-NEXT:    v_readlane_b32 s55, v32, 11
; GFX9-NEXT:    v_readlane_b32 s54, v32, 10
; GFX9-NEXT:    v_readlane_b32 s53, v32, 9
; GFX9-NEXT:    v_readlane_b32 s52, v32, 8
; GFX9-NEXT:    v_readlane_b32 s51, v32, 7
; GFX9-NEXT:    v_readlane_b32 s50, v32, 6
; GFX9-NEXT:    v_readlane_b32 s49, v32, 5
; GFX9-NEXT:    v_readlane_b32 s48, v32, 4
; GFX9-NEXT:    v_readlane_b32 s39, v32, 3
; GFX9-NEXT:    v_readlane_b32 s38, v32, 2
; GFX9-NEXT:    v_readlane_b32 s37, v32, 1
; GFX9-NEXT:    v_readlane_b32 s36, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v16f64_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s36, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s63, v9
; GFX11-NEXT:    v_readfirstlane_b32 s62, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s61, v7
; GFX11-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-NEXT:    v_readfirstlane_b32 s59, v5
; GFX11-NEXT:    v_readfirstlane_b32 s58, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s57, v3
; GFX11-NEXT:    v_readfirstlane_b32 s56, v2
; GFX11-NEXT:    s_mov_b32 s47, s23
; GFX11-NEXT:    s_mov_b32 s46, s22
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    s_mov_b32 s45, s21
; GFX11-NEXT:    s_mov_b32 s44, s20
; GFX11-NEXT:    s_mov_b32 s43, s19
; GFX11-NEXT:    s_mov_b32 s42, s18
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    s_mov_b32 s48, s24
; GFX11-NEXT:    s_mov_b32 s41, s17
; GFX11-NEXT:    s_mov_b32 s40, s16
; GFX11-NEXT:    s_mov_b32 s39, s3
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    s_mov_b32 s49, s25
; GFX11-NEXT:    s_mov_b32 s38, s2
; GFX11-NEXT:    s_mov_b32 s37, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    s_mov_b32 s50, s26
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    s_mov_b32 s51, s27
; GFX11-NEXT:    v_writelane_b32 v32, s52, 8
; GFX11-NEXT:    s_mov_b32 s52, s28
; GFX11-NEXT:    v_writelane_b32 v32, s53, 9
; GFX11-NEXT:    s_mov_b32 s53, s29
; GFX11-NEXT:    v_writelane_b32 v32, s54, 10
; GFX11-NEXT:    v_readfirstlane_b32 s54, v0
; GFX11-NEXT:    v_writelane_b32 v32, s55, 11
; GFX11-NEXT:    v_readfirstlane_b32 s55, v1
; GFX11-NEXT:    v_writelane_b32 v32, s64, 12
; GFX11-NEXT:    v_readfirstlane_b32 s64, v10
; GFX11-NEXT:    v_writelane_b32 v32, s65, 13
; GFX11-NEXT:    v_readfirstlane_b32 s65, v11
; GFX11-NEXT:    v_writelane_b32 v32, s66, 14
; GFX11-NEXT:    v_readfirstlane_b32 s66, v12
; GFX11-NEXT:    v_writelane_b32 v32, s67, 15
; GFX11-NEXT:    v_readfirstlane_b32 s67, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB87_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB87_4
; GFX11-NEXT:  .LBB87_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, s51, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, s50, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, s49, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, s48, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, s47, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, s46, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, s45, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, s44, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, s43, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, s42, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, s41, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, s40, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, s39, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, s38, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, s37, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, s36, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v31, s67, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, s66, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, s65, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, s64, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, s63, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, s62, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, s61, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, s60, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, s59, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, s58, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, s57, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, s56, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, s55, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, s54, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, s53, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, s52, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_branch .LBB87_5
; GFX11-NEXT:  .LBB87_3:
; GFX11-NEXT:    s_branch .LBB87_2
; GFX11-NEXT:  .LBB87_4:
; GFX11-NEXT:    v_dual_mov_b32 v0, s36 :: v_dual_mov_b32 v1, s37
; GFX11-NEXT:    v_dual_mov_b32 v2, s38 :: v_dual_mov_b32 v3, s39
; GFX11-NEXT:    v_dual_mov_b32 v4, s40 :: v_dual_mov_b32 v5, s41
; GFX11-NEXT:    v_dual_mov_b32 v6, s42 :: v_dual_mov_b32 v7, s43
; GFX11-NEXT:    v_dual_mov_b32 v8, s44 :: v_dual_mov_b32 v9, s45
; GFX11-NEXT:    v_dual_mov_b32 v10, s46 :: v_dual_mov_b32 v11, s47
; GFX11-NEXT:    v_dual_mov_b32 v12, s48 :: v_dual_mov_b32 v13, s49
; GFX11-NEXT:    v_dual_mov_b32 v14, s50 :: v_dual_mov_b32 v15, s51
; GFX11-NEXT:    v_dual_mov_b32 v16, s52 :: v_dual_mov_b32 v17, s53
; GFX11-NEXT:    v_dual_mov_b32 v18, s54 :: v_dual_mov_b32 v19, s55
; GFX11-NEXT:    v_dual_mov_b32 v20, s56 :: v_dual_mov_b32 v21, s57
; GFX11-NEXT:    v_dual_mov_b32 v22, s58 :: v_dual_mov_b32 v23, s59
; GFX11-NEXT:    v_dual_mov_b32 v24, s60 :: v_dual_mov_b32 v25, s61
; GFX11-NEXT:    v_dual_mov_b32 v26, s62 :: v_dual_mov_b32 v27, s63
; GFX11-NEXT:    v_dual_mov_b32 v28, s64 :: v_dual_mov_b32 v29, s65
; GFX11-NEXT:    v_dual_mov_b32 v30, s66 :: v_dual_mov_b32 v31, s67
; GFX11-NEXT:  .LBB87_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s67, v32, 15
; GFX11-NEXT:    v_readlane_b32 s66, v32, 14
; GFX11-NEXT:    v_readlane_b32 s65, v32, 13
; GFX11-NEXT:    v_readlane_b32 s64, v32, 12
; GFX11-NEXT:    v_readlane_b32 s55, v32, 11
; GFX11-NEXT:    v_readlane_b32 s54, v32, 10
; GFX11-NEXT:    v_readlane_b32 s53, v32, 9
; GFX11-NEXT:    v_readlane_b32 s52, v32, 8
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <16 x double>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <16 x double>
  br label %end

end:
  %phi = phi <16 x double> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <16 x double> %phi
}

define <64 x bfloat> @bitcast_v128i8_to_v64bf16(<128 x i8> %a, i32 %b) {
; SI-LABEL: bitcast_v128i8_to_v64bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:388
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:108
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v5
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v7
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v11
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v15
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v19
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v21
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v23
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v27
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v29
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v17
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v18
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:136
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; kill: killed $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:168
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v25
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:196
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:212
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v18, 8, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:232
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v1
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:252
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v4
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:300
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:292
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:308
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v54, 24, v5
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v43, 8, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:336
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:344
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:316
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:332
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:340
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v45, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v47, 24, v13
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v59, 8, v14
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:352
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:368
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:376
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:348
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:364
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:360
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:356
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:372
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v60, 24, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:384
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:380
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v61, 24, v13
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; kill: killed $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v14
; SI-NEXT:    ; kill: killed $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; kill: killed $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:28
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:20
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:100
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:8
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:92
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:116
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:84
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:76
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:72
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:68
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:60
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:52
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:44
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; kill: killed $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB88_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v31, v7, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v6
; SI-NEXT:    v_or_b32_e32 v32, v4, v2
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v29, v45, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    v_or_b32_e32 v33, v0, v4
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_or_b32_e32 v34, v47, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v41
; SI-NEXT:    v_or_b32_e32 v0, v0, v59
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v8
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v27, v54, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v9
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v2, v2, v43
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v13, 24, v13
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v42
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v35, v60, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v56
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v46
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v36, v5, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v44
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v37, v61, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v58
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v57
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v4, v4, v63
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v38, v62, v5
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v4
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v14, 24, v14
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v15, 24, v15
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v21
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v39, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v48, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v49, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v49
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v50, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v51, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v52, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v52
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v53, v15, v19
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v19, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 24, v19
; SI-NEXT:    v_or_b32_e32 v26, v19, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v48
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v51
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v25, v3, v15
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v24, v17, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v3, v3, v18
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v3
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 24, v17
; SI-NEXT:    v_or_b32_e32 v30, v17, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v28, v10, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v50
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v53
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:  .LBB88_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB88_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v4
; SI-NEXT:    s_movk_i32 s6, 0x300
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_or_b32_e32 v2, v7, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v7, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_or_b32_e32 v5, v47, v5
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v5, v0
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v9
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v11
; SI-NEXT:    v_or_b32_e32 v5, v43, v5
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_or_b32_e32 v9, v45, v9
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v9, v5
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    v_or_b32_e32 v8, v54, v8
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v8, v6
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v58
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v57
; SI-NEXT:    v_or_b32_e32 v12, v63, v12
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, 0x300, v12
; SI-NEXT:    v_or_b32_e32 v13, v62, v13
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v56
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v14, 8, v46
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v44
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_or_b32_e32 v14, v61, v14
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v41
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v42
; SI-NEXT:    v_or_b32_e32 v14, v59, v14
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v15, v60, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    s_mov_b32 s7, 0x3000000
; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v7
; SI-NEXT:    v_add_i32_e32 v0, vcc, s7, v0
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v6
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v4
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v0
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v10, v2
; SI-NEXT:    v_or_b32_e32 v8, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v8
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_or_b32_e32 v1, v18, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_or_b32_e32 v2, v17, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v9, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v9
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v8
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v10, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v10
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v11, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v10, vcc, s7, v11
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v10
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v15, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v15, 8, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v16, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v16
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v17, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s7, v17
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v18, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s7, v18
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v18
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v19, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v19
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v21, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v21
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v23, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v25, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v25
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v26, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v25, vcc, s7, v26
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v27, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v28, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v26, vcc, s7, v28
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v3
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v29, v2, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    v_or_b32_e32 v30, v2, v1
; SI-NEXT:    v_add_i32_e32 v2, vcc, s7, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v27
; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s7, v12
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v30, v12
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v12
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v2
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s7, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v23, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v23, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 8, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v29, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v30, v19
; SI-NEXT:    v_or_b32_e32 v15, v19, v15
; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 8, v30
; SI-NEXT:    v_or_b32_e32 v15, v30, v15
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_or_b32_e32 v15, v30, v15
; SI-NEXT:    v_add_i32_e32 v30, vcc, s7, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v31, v15
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_or_b32_e32 v15, v31, v15
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 8, v32
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v31, vcc, s6, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v32
; SI-NEXT:    v_and_b32_e32 v32, 0xff, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v31, vcc, s7, v31
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v15
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v19
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v23
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v14
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v12
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v13
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v21
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v10
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v9
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v8
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v7
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v6
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v5
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v0
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
; SI-NEXT:  .LBB88_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v13
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v15
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v23
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v14
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v39
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v49
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v52
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_alignbit_b32 v18, v18, v19, 16
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_alignbit_b32 v19, v19, v20, 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v22
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v50
; SI-NEXT:    v_alignbit_b32 v21, v21, v22, 16
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_alignbit_b32 v22, v22, v23, 16
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v24
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v51
; SI-NEXT:    v_alignbit_b32 v23, v23, v24, 16
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v28
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v30
; SI-NEXT:    v_alignbit_b32 v24, v24, v25, 16
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_alignbit_b32 v25, v25, v26, 16
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v40
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_alignbit_b32 v26, v26, v27, 16
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v53
; SI-NEXT:    v_alignbit_b32 v27, v27, v28, 16
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v34
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v33
; SI-NEXT:    v_alignbit_b32 v28, v28, v29, 16
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v35
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v55
; SI-NEXT:    v_alignbit_b32 v29, v29, v30, 16
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v37
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v36
; SI-NEXT:    v_alignbit_b32 v30, v30, v31, 16
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v38
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v64bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:4
; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:376
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:120
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB88_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s6, 0xc0c0004
; VI-NEXT:    v_perm_b32 v10, v37, v36, s6
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v11, v59, v34, s6
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_perm_b32 v12, v57, v32, s6
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v47, v46, s6
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v43, v42, s6
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v55, v54, s6
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v16, v51, v50, s6
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v1, v0, s6
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s6
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v9, s6
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_perm_b32 v9, v39, v62, s6
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_perm_b32 v8, v38, v61, s6
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_perm_b32 v9, v35, v60, s6
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_perm_b32 v10, v33, v58, s6
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_perm_b32 v11, v63, v56, s6
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v45, v44, s6
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_perm_b32 v13, v41, v40, s6
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_perm_b32 v14, v53, v52, s6
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v49, v15, s6
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v17, v18, v19
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v20, v21, s6
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v22, s6
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v27, v28, s6
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v29, v30, s6
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v30, v30, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v31, v31, v48, s6
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v32, v33, v32, s6
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:  .LBB88_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB88_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v18, 0x300
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v61
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v60
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v58
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v56
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v44
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v40
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v52
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    s_waitcnt vmcnt(11)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    s_waitcnt vmcnt(10)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    s_waitcnt vmcnt(9)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    s_waitcnt vmcnt(8)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    s_waitcnt vmcnt(7)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    s_waitcnt vmcnt(6)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v9
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v1, 0x300, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v3
; VI-NEXT:    v_or_b32_e32 v0, v2, v0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v1, v1, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
; VI-NEXT:    v_or_b32_e32 v2, v3, v2
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
; VI-NEXT:    v_or_b32_e32 v3, v4, v3
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
; VI-NEXT:    v_or_b32_e32 v4, v5, v4
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v5, v5, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
; VI-NEXT:    v_or_b32_e32 v5, v6, v5
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
; VI-NEXT:    v_or_b32_e32 v6, v7, v6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
; VI-NEXT:    v_or_b32_e32 v7, v8, v7
; VI-NEXT:    v_add_u16_e32 v8, 3, v39
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v62
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 3, v38
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
; VI-NEXT:    v_or_b32_e32 v8, v9, v8
; VI-NEXT:    v_add_u16_e32 v9, 3, v37
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v36
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 3, v35
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
; VI-NEXT:    v_or_b32_e32 v9, v10, v9
; VI-NEXT:    v_add_u16_e32 v10, 3, v59
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v34
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 3, v33
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
; VI-NEXT:    v_or_b32_e32 v10, v11, v10
; VI-NEXT:    v_add_u16_e32 v11, 3, v57
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v32
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 3, v63
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
; VI-NEXT:    v_or_b32_e32 v11, v12, v11
; VI-NEXT:    v_add_u16_e32 v12, 3, v47
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v46
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 3, v45
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    v_add_u16_e32 v13, 3, v43
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v42
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 3, v41
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_add_u16_e32 v14, 3, v55
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v54
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v15, 3, v53
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_add_u16_e32 v15, 3, v51
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v50
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v16, 3, v49
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v21, v16, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v19, 0x300, v20
; VI-NEXT:    v_or_b32_e32 v16, v19, v16
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v17, v17, v21
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v48
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v31, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v31, 0x300, v31
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v30, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v33, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v29, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v35, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v28, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v35
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v37, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v27, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v37
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v38, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v26, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v38
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v39, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v39, v39, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v25, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v39
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v48, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v48, v48, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v24, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v48
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v49, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v49, v49, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v23, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v49
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v50, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v50, v50, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v22, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v50
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v51, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v51, v51, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    v_or_b32_sdwa v21, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v51
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    v_or_b32_sdwa v52, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v36, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v36, 0x300, v36
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v53, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v54, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v55, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v32, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v34, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v20, v52, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v19, v53, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v18, v55, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v36, v20
; VI-NEXT:    v_or_b32_e32 v30, v30, v34
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v52, 3, v52
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v53
; VI-NEXT:    v_or_b32_sdwa v52, v53, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v52, 0x300, v52
; VI-NEXT:    v_or_b32_e32 v18, v52, v18
; VI-NEXT:    v_add_u16_e32 v52, 0x300, v54
; VI-NEXT:    v_or_b32_e32 v19, v52, v19
; VI-NEXT:  .LBB88_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v64bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:264
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:272
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:288
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:296
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:304
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:320
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
; GFX9-NEXT:    s_waitcnt vmcnt(27)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-NEXT:    s_waitcnt vmcnt(26)
; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v1
; GFX9-NEXT:    s_waitcnt vmcnt(25)
; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(24)
; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(23)
; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v4
; GFX9-NEXT:    s_waitcnt vmcnt(22)
; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v6
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:100
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:92
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v7
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v9
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v11
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v13
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v15
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v17
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:352
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:348
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:344
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:360
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:384
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:356
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:364
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:380
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:340
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:12
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:372
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:376
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB88_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0xc0c0004
; GFX9-NEXT:    s_mov_b32 s7, 0x5040100
; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v10, v47, v45, s6
; GFX9-NEXT:    v_or_b32_sdwa v11, v40, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v12, v55, v54, s6
; GFX9-NEXT:    v_perm_b32 v17, v43, v42, s6
; GFX9-NEXT:    v_perm_b32 v18, v58, v57, s6
; GFX9-NEXT:    v_perm_b32 v19, v63, v46, s6
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_perm_b32 v1, v3, v2, s7
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s7
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s7
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s7
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s7
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s7
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v33, s6
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s7
; GFX9-NEXT:    v_or_b32_sdwa v8, v60, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s7
; GFX9-NEXT:    v_or_b32_sdwa v9, v44, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v9, v9, v10, s7
; GFX9-NEXT:    v_or_b32_sdwa v10, v41, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s7
; GFX9-NEXT:    v_or_b32_sdwa v11, v52, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v11, v11, v12, s7
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v12, v50, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s7
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v14, v37, v36, s6
; GFX9-NEXT:    v_perm_b32 v13, v13, v14, s7
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s7
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_or_b32_sdwa v15, v62, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v35, s6
; GFX9-NEXT:    v_perm_b32 v15, v15, v16, s7
; GFX9-NEXT:    v_perm_b32 v16, v32, v56, s6
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s7
; GFX9-NEXT:    v_perm_b32 v17, v19, v18, s7
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s6
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s7
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s7
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v20, v20, v61, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s7
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s7
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s7
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s7
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s7
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s7
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s7
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s7
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s7
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s7
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s7
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v32, v33, v32, s6
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s7
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:  .LBB88_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB88_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v33
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v45
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v40
; GFX9-NEXT:    v_or_b32_sdwa v11, v38, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v54
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v46
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v56
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v42
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(19)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v1, 0x300, v1
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v3
; GFX9-NEXT:    v_perm_b32 v0, v2, v0, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v1, v1, v4, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v60
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
; GFX9-NEXT:    v_or_b32_sdwa v8, v53, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_or_b32_sdwa v9, v51, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v47
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v44
; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v41
; GFX9-NEXT:    v_or_b32_sdwa v10, v39, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v55
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v52
; GFX9-NEXT:    v_or_b32_sdwa v12, v34, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v50
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v48
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v37
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v36
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v14
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v14
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v35
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v62
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v58
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v57
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v63
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v32
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v43
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v16
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v18
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v19
; GFX9-NEXT:    v_perm_b32 v16, v18, v16, s6
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
; GFX9-NEXT:    v_perm_b32 v17, v17, v20, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v32, 0x300, v19
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v33, 0x300, v21
; GFX9-NEXT:    v_perm_b32 v30, v33, v30, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v34, 0x300, v23
; GFX9-NEXT:    v_perm_b32 v29, v34, v29, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v35, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v36, 0x300, v36
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v39
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v39, 0x300, v50
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v51
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v53
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v54, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v55
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v61
; GFX9-NEXT:    v_add_u16_e32 v50, 0x300, v40
; GFX9-NEXT:    v_perm_b32 v21, v50, v21, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v41
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v51, 0x300, v42
; GFX9-NEXT:    v_perm_b32 v20, v51, v20, s6
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v43
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v44, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v45, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v45
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v46, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v24
; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v26
; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v37
; GFX9-NEXT:    v_add_u16_e32 v37, 0x300, v38
; GFX9-NEXT:    v_add_u16_e32 v38, 0x300, v48
; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v49
; GFX9-NEXT:    v_add_u16_e32 v48, 0x300, v52
; GFX9-NEXT:    v_add_u16_e32 v49, 0x300, v54
; GFX9-NEXT:    v_add_u16_e32 v52, 0x300, v44
; GFX9-NEXT:    v_add_u16_e32 v53, 0x300, v46
; GFX9-NEXT:    v_perm_b32 v18, v53, v18, s6
; GFX9-NEXT:    v_perm_b32 v19, v52, v19, s6
; GFX9-NEXT:    v_perm_b32 v22, v49, v22, s6
; GFX9-NEXT:    v_perm_b32 v23, v48, v23, s6
; GFX9-NEXT:    v_perm_b32 v24, v39, v24, s6
; GFX9-NEXT:    v_perm_b32 v25, v38, v25, s6
; GFX9-NEXT:    v_perm_b32 v26, v37, v26, s6
; GFX9-NEXT:    v_perm_b32 v27, v36, v27, s6
; GFX9-NEXT:    v_perm_b32 v28, v35, v28, s6
; GFX9-NEXT:  .LBB88_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64bf16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:392
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:384
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:380
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:376
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:372
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v31, off, s32 offset:368
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:364
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v70, off, s32 offset:360
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v39, off, s32 offset:356
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:352
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:348
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:344
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:340
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:336
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v35, off, s32 offset:332
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v69, off, s32 offset:328
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v38, off, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v68, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v36, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v67, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v66, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v32, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v33, off, s32 offset:260
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:132
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v65, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:388
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v52, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v53, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v54, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v55, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v64, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v71, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v51, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v50, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v48, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v49, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v37, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v34, off, s32 offset:12
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.l, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.l, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.l, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, v0.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v69.l, 8, v69.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v70.l, 8, v70.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v69.h, 8, v69.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v68.h, 8, v68.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.l, 8, v67.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v68.l, 8, v68.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.h, 8, v67.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.h, 8, v66.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.l, 8, v66.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(33)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v65.h, 8, v65.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(29)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v88
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.l, 8, v52.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.h, 8, v52.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(25)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v53.l, 8, v53.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(24)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.l, 8, v53.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v53.h, 8, v54.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(22)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.l, 8, v55.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.h, 8, v55.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(20)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.h, 8, v64.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v65.l, 8, v64.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(18)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v64.l, 8, v71.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v64.h, 8, v71.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.l, 8, v70.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v70.h, 8, v31.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.h, 8, v31.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB88_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v130, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v165, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v45, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v60, v41, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v37.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v160, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v48.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v3.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v177, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v13.h, 0xff, v51.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v40, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v32.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v7.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v58, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v9.l, v53.l
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v49.h
; GFX11-TRUE16-NEXT:    v_and_b16 v11.l, 0xff, v38.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v11.h, v55.l
; GFX11-TRUE16-NEXT:    v_and_b16 v12.h, 0xff, v50.h
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v49.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v13.h, v65.l
; GFX11-TRUE16-NEXT:    v_and_b16 v15.l, 0xff, v51.h
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v50.l
; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v14.h, v65.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v63, v56, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v98, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v10.h, v54.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v53.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v77, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v102, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v12.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.h, v55.h
; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v13.l, v54.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v79, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v119, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v14.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v15.l, v64.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v15.h, v64.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v83, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v144, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v16.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v96, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v149, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v17.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v99, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v163, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v18.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v103, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v179, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v117, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v19.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v148, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v44, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v25.l, 0xff, v34.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v146, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v20.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v145, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v35.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v181, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v21.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v161, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v39.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v57, v182, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v22.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v43, v167, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v48.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v113, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v23.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v180, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v2.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v147, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v26.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v25.l, v67.h
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v36.h
; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v33.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v4.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v164, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v27.h, v68.h
; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v38.h
; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v35.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v62, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v29.h, v69.l
; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v39.h
; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v36.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v37.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.h, 0xff, v34.h
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v33.h
; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v32.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v30.l, v71.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v59, v183, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v26.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v26.h, v68.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v27.l, v67.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v72, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v28.l
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.h, v69.h
; GFX11-TRUE16-NEXT:    v_or_b16 v28.h, v29.l, v70.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v74, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v30.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v31.l, v71.l
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v31.h, v70.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v78, v75, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.l, v52.l
; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v8.h, v52.h
; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.l, v66.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v24.h, v66.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v32.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_hi16
; GFX11-TRUE16-NEXT:  .LBB88_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB88_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v78.l, 3
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v48.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v39.h, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v75.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v36.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v74.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v43.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v71.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v71.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v39.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v70.h, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.l, 8, v46.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v38.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v35.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.l, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v69.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v72.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v69.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v70.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v61.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v35.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v36.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v33.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v59.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v68.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v68.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v34.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v183.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v67.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.h, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v33.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v32.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v67.h, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v180.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v44.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v66.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v151.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v66.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v176.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v167.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v179.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v161.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v150.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v134.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v163.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v145.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v131.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v118.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v149.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v129.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v148.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v144.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v115.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v116.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v103.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v119.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v86.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v112.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v99.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v85.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v102.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v96.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v84.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v82.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v98.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v83.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v81.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v32.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v80.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v3.l, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v51.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v50.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v65.h, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v79.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v51.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v64.l, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v64.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v76.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v50.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v65.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v49.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v77.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v55.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v48.h, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v73.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v49.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v54.h, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v38.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v55.l, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v54.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v63.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v37.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v53.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v56.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v37.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v34.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v53.l, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v58.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v52.l, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v52.h, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v42.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v60.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v62.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v41.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v47.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v57.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v40.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v182.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v166.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v45.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.h, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v178.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v164.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v181.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v133.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v162.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v177.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v165.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v147.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v132.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v135.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v128.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v3.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v3.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v146.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v160.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v113.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v130.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v32.h, v117.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.l, 8, v114.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.l, 8, v101.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.l, 8, v97.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v87.l
; GFX11-TRUE16-NEXT:    v_and_b16 v32.h, 0xff, v32.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.l, 8, v100.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.l, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v32.l, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v33.l, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v33.h, v2.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.h, v34.l, v32.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v32.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v32.h
; GFX11-TRUE16-NEXT:  .LBB88_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:392
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:488
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64bf16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, v24 :: v_dual_mov_b32 v51, v19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, v23 :: v_dual_mov_b32 v71, v20
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, v22 :: v_dual_mov_b32 v67, v17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v68, v21 :: v_dual_mov_b32 v65, v18
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, v16 :: v_dual_mov_b32 v50, v15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v14 :: v_dual_mov_b32 v53, v9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v13 :: v_dual_mov_b32 v49, v10
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, v12 :: v_dual_mov_b32 v33, v7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v55, v8
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v39, v6 :: v_dual_mov_b32 v36, v5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, v4 :: v_dual_mov_b32 v37, v2
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v38, v0
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v0, off, s32 offset:384
; GFX11-FAKE16-NEXT:    scratch_load_u16 v105, off, s32 offset:380
; GFX11-FAKE16-NEXT:    scratch_load_u16 v93, off, s32 offset:376
; GFX11-FAKE16-NEXT:    scratch_load_u16 v104, off, s32 offset:372
; GFX11-FAKE16-NEXT:    scratch_load_u16 v1, off, s32 offset:368
; GFX11-FAKE16-NEXT:    scratch_load_u16 v94, off, s32 offset:364
; GFX11-FAKE16-NEXT:    scratch_load_u16 v2, off, s32 offset:360
; GFX11-FAKE16-NEXT:    scratch_load_u16 v106, off, s32 offset:356
; GFX11-FAKE16-NEXT:    scratch_load_u16 v3, off, s32 offset:352
; GFX11-FAKE16-NEXT:    scratch_load_u16 v91, off, s32 offset:348
; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:344
; GFX11-FAKE16-NEXT:    scratch_load_u16 v89, off, s32 offset:340
; GFX11-FAKE16-NEXT:    scratch_load_u16 v4, off, s32 offset:336
; GFX11-FAKE16-NEXT:    scratch_load_u16 v88, off, s32 offset:332
; GFX11-FAKE16-NEXT:    scratch_load_u16 v5, off, s32 offset:328
; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_load_u16 v6, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v62, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v7, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v73, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v8, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v79, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v9, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v63, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v45, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v61, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v10, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v11, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v72, off, s32 offset:260
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v57, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:188
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:132
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v12, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_b32 v13, off, s32 offset:388
; GFX11-FAKE16-NEXT:    scratch_load_u16 v14, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v15, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v16, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v17, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v18, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v19, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v20, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v21, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v22, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v23, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v24, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:12
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v83, v30 :: v_dual_mov_b32 v86, v28
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, v29 :: v_dual_mov_b32 v80, v27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, v26 :: v_dual_mov_b32 v81, v25
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(62)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v127, 8, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v126, 8, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v125, 8, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v123, 8, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v122, 8, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v124, 8, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v121, 8, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v120, 8, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v111, 8, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v110, 8, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v108, 8, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v109, 8, v11
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(33)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v107, 8, v12
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(29)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(28)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v182, 8, v14
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(27)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v166, 8, v15
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(26)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v180, 8, v16
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(25)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v46, 8, v17
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(24)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v47, 8, v18
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(23)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v56, 8, v19
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(22)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v75, 8, v20
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(21)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v59, 8, v21
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(20)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v74, 8, v22
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(19)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v95, 8, v23
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(18)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v90, 8, v24
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB88_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v38, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v37, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v39, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v64, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v52, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v69, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v3, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v55, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v49, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v65, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v84, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v82, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v71, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v70, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v3, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v5, v4, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v7, v6, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v11, v10, 0x5040100
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v118
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v113
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v9, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v86, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v83, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v99
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v7, v182
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v134
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v129
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v180
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v112, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v166
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v7, v46
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v47
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v9, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v132
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v11, v14, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v151
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v10, v12, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v13, v15, 0x5040100
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v56
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v145
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v150
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v11, v75
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v167
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v162
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v130, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v59
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v74
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v148, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v11, v95
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v90
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v116
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v12, v13, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v14, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v15, v18, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v17, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v20, v107
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v115, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v103, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v98, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v131, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v117, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v147, v146, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v133, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v164, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v149, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v18, v17, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v20, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v21, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v24, v23, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v177, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v165, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v183, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v179, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v43, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v40, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v60, v57, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v44, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v72
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v21, v20, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v23, v22, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v25, v24, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v58
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v28, v109
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v79
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v73
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v63
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v108
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v61, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v111
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v120
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v110
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v91
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v28, v27, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v88
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v92
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v94
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v105
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v27, v122
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v106
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v30, v121
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v76, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v124
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v31, v123
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v89, v77, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v27, v125
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v126
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v34, v127
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v104, v93, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v28, v30, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v31, v35, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v33, v36, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v34, v37, 0x5040100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr89
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr91
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr106
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr93
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr105
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr95
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr107
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr109
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr108
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr111
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr120
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr121
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr124
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr122
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr125
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr126
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr127
; GFX11-FAKE16-NEXT:  .LBB88_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB88_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v104, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v105, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v93
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v106, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v94, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v92, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v127, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v89, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v125, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v126, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v89, 0x300, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v1, 8, v77
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v91, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v77, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v88, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v1, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v76, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v62
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v123, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v124, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v78, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v62, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v122, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v5, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v79, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v73, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v73, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v121, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v61, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v72, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v61, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v63, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v120, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v111, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v45
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v45, 0x300, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v110, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v58, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v60, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v42
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v58, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v109, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v57
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v44, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v43, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v108, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v41
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v40, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v40, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v183, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v181, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v179, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v178
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v176
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v177, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v176, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v165, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v163
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v161
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v164, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v161, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v160
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v149, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v147, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v147, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v144
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v133, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v131, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v119
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v128
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v117, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v117, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v103, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v102, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v98, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v97
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v115, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v96, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v116, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(16)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v167, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v97, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v107, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v162, 3
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(13)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v148, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v98, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v95, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v150, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v135
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(11)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v151, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v90, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v145, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v100, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v74, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v75, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(8)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v130, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v103, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v132, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v59, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(6)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v134, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v129, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v114, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v56, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v112, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v112, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v46, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v47, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v113, 3
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v118, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v101, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v99, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v86, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v180, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v182, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v83, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v83, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v166, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v87
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v84, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v84, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v82, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v81
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v80
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v80, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v71, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v70, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v68
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v4, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v69, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v68, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v5, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v66, 8, v67
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v6, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v65, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v64, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v54
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v52, v52, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v67, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v66, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v51, 8, v51
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v50, 8, v50
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v55, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v51, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v49, v49, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v51, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v50, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v48, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v39, v39, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v38, v38, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v37, v37, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v53, 8, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xff, v49
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v36, 8, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v35, 8, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v33, 8, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v53, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v36, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v35, v38
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v34, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v48
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v37, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v34, v35, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v33, v3, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v32, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v36, v51, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v37, v4, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v6, v5, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v68, v67, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v80, v7, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v84, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v83, v9, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v101, v10, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v112, v11, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v114, v12, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v103, v13, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v100, v14, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v98, v15, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v97, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v96, v17, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v102, v18, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v117, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v147, v20, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v161, v21, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v176, v22, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v181, v23, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v40, v24, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v58, v25, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v45, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v61, v27, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v73, v28, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v62, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v77, v30, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v89, v31, 0x5040100
; GFX11-FAKE16-NEXT:  .LBB88_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define inreg <64 x bfloat> @bitcast_v128i8_to_v64bf16_scalar(<128 x i8> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v128i8_to_v64bf16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    ; implicit-def: $vgpr43 : SGPR spill to VGPR lane
; SI-NEXT:    v_readfirstlane_b32 s4, v30
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v43, s29, 0
; SI-NEXT:    v_writelane_b32 v43, s28, 1
; SI-NEXT:    v_writelane_b32 v43, s27, 2
; SI-NEXT:    v_writelane_b32 v43, s26, 3
; SI-NEXT:    v_writelane_b32 v43, s25, 4
; SI-NEXT:    v_writelane_b32 v43, s24, 5
; SI-NEXT:    v_writelane_b32 v43, s23, 6
; SI-NEXT:    v_writelane_b32 v43, s22, 7
; SI-NEXT:    v_writelane_b32 v43, s4, 8
; SI-NEXT:    v_readfirstlane_b32 s4, v29
; SI-NEXT:    v_writelane_b32 v43, s4, 9
; SI-NEXT:    v_readfirstlane_b32 s4, v28
; SI-NEXT:    v_writelane_b32 v40, s30, 0
; SI-NEXT:    v_writelane_b32 v43, s4, 10
; SI-NEXT:    v_readfirstlane_b32 s4, v27
; SI-NEXT:    v_writelane_b32 v40, s31, 1
; SI-NEXT:    v_writelane_b32 v43, s4, 11
; SI-NEXT:    v_readfirstlane_b32 s4, v26
; SI-NEXT:    v_writelane_b32 v40, s34, 2
; SI-NEXT:    v_writelane_b32 v43, s4, 12
; SI-NEXT:    v_readfirstlane_b32 s4, v25
; SI-NEXT:    v_writelane_b32 v40, s35, 3
; SI-NEXT:    v_writelane_b32 v43, s4, 13
; SI-NEXT:    v_readfirstlane_b32 s4, v24
; SI-NEXT:    v_writelane_b32 v40, s36, 4
; SI-NEXT:    v_writelane_b32 v43, s4, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v23
; SI-NEXT:    v_writelane_b32 v40, s37, 5
; SI-NEXT:    v_writelane_b32 v43, s4, 15
; SI-NEXT:    v_readfirstlane_b32 s4, v22
; SI-NEXT:    v_writelane_b32 v40, s38, 6
; SI-NEXT:    v_writelane_b32 v43, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v21
; SI-NEXT:    v_writelane_b32 v40, s39, 7
; SI-NEXT:    v_writelane_b32 v43, s4, 17
; SI-NEXT:    v_readfirstlane_b32 s4, v20
; SI-NEXT:    v_writelane_b32 v40, s48, 8
; SI-NEXT:    v_writelane_b32 v43, s4, 18
; SI-NEXT:    v_readfirstlane_b32 s4, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:316
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:308
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:300
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:292
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:268
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:260
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:252
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:244
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:332
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:236
; SI-NEXT:    v_writelane_b32 v40, s49, 9
; SI-NEXT:    v_writelane_b32 v40, s50, 10
; SI-NEXT:    v_writelane_b32 v40, s51, 11
; SI-NEXT:    v_writelane_b32 v40, s52, 12
; SI-NEXT:    s_mov_b32 s37, s18
; SI-NEXT:    v_readfirstlane_b32 s18, v18
; SI-NEXT:    v_readfirstlane_b32 s38, v17
; SI-NEXT:    v_readfirstlane_b32 s39, v16
; SI-NEXT:    v_readfirstlane_b32 s95, v15
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:228
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:220
; SI-NEXT:    v_writelane_b32 v40, s53, 13
; SI-NEXT:    v_writelane_b32 v40, s54, 14
; SI-NEXT:    v_writelane_b32 v40, s55, 15
; SI-NEXT:    v_writelane_b32 v40, s64, 16
; SI-NEXT:    s_mov_b32 s73, s19
; SI-NEXT:    v_readfirstlane_b32 s55, v14
; SI-NEXT:    v_readfirstlane_b32 s34, v13
; SI-NEXT:    v_readfirstlane_b32 s64, v12
; SI-NEXT:    v_readfirstlane_b32 s19, v11
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204
; SI-NEXT:    v_readfirstlane_b32 s54, v10
; SI-NEXT:    v_readfirstlane_b32 s78, v9
; SI-NEXT:    v_readfirstlane_b32 s88, v8
; SI-NEXT:    v_readfirstlane_b32 s49, v7
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:196
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:188
; SI-NEXT:    v_readfirstlane_b32 s52, v2
; SI-NEXT:    v_readfirstlane_b32 s7, v1
; SI-NEXT:    v_readfirstlane_b32 s75, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176
; SI-NEXT:    v_writelane_b32 v43, s4, 19
; SI-NEXT:    v_readfirstlane_b32 s53, v6
; SI-NEXT:    v_readfirstlane_b32 s76, v5
; SI-NEXT:    v_readfirstlane_b32 s77, v4
; SI-NEXT:    v_readfirstlane_b32 s48, v3
; SI-NEXT:    v_writelane_b32 v40, s65, 17
; SI-NEXT:    v_writelane_b32 v40, s66, 18
; SI-NEXT:    v_writelane_b32 v40, s67, 19
; SI-NEXT:    v_writelane_b32 v40, s68, 20
; SI-NEXT:    ; implicit-def: $vgpr42 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v40, s69, 21
; SI-NEXT:    v_writelane_b32 v40, s70, 22
; SI-NEXT:    v_writelane_b32 v40, s71, 23
; SI-NEXT:    v_writelane_b32 v40, s80, 24
; SI-NEXT:    v_writelane_b32 v40, s81, 25
; SI-NEXT:    v_writelane_b32 v40, s82, 26
; SI-NEXT:    v_writelane_b32 v40, s83, 27
; SI-NEXT:    v_writelane_b32 v40, s84, 28
; SI-NEXT:    v_writelane_b32 v40, s85, 29
; SI-NEXT:    v_writelane_b32 v40, s86, 30
; SI-NEXT:    v_writelane_b32 v40, s87, 31
; SI-NEXT:    v_writelane_b32 v40, s96, 32
; SI-NEXT:    v_writelane_b32 v40, s97, 33
; SI-NEXT:    v_writelane_b32 v40, s98, 34
; SI-NEXT:    v_writelane_b32 v40, s99, 35
; SI-NEXT:    s_mov_b32 s6, s20
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s91, v19
; SI-NEXT:    v_readfirstlane_b32 s43, v20
; SI-NEXT:    v_readfirstlane_b32 s47, v21
; SI-NEXT:    v_readfirstlane_b32 s56, v22
; SI-NEXT:    v_readfirstlane_b32 s59, v23
; SI-NEXT:    v_readfirstlane_b32 s44, v24
; SI-NEXT:    v_readfirstlane_b32 s60, v25
; SI-NEXT:    v_readfirstlane_b32 s62, v26
; SI-NEXT:    v_readfirstlane_b32 s63, v27
; SI-NEXT:    v_readfirstlane_b32 s72, v28
; SI-NEXT:    v_readfirstlane_b32 s79, v29
; SI-NEXT:    v_readfirstlane_b32 s89, v30
; SI-NEXT:    v_readfirstlane_b32 s22, v31
; SI-NEXT:    v_readfirstlane_b32 s50, v32
; SI-NEXT:    v_readfirstlane_b32 s36, v33
; SI-NEXT:    v_readfirstlane_b32 s14, v34
; SI-NEXT:    v_readfirstlane_b32 s93, v35
; SI-NEXT:    v_readfirstlane_b32 s92, v36
; SI-NEXT:    v_readfirstlane_b32 s30, v37
; SI-NEXT:    v_readfirstlane_b32 s51, v38
; SI-NEXT:    v_readfirstlane_b32 s58, v39
; SI-NEXT:    v_readfirstlane_b32 s40, v48
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v49
; SI-NEXT:    v_readfirstlane_b32 s61, v50
; SI-NEXT:    v_readfirstlane_b32 s4, v51
; SI-NEXT:    v_writelane_b32 v43, s4, 20
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:172
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:148
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:132
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:124
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:116
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v43, s4, 21
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s4, v17
; SI-NEXT:    v_writelane_b32 v43, s4, 22
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_writelane_b32 v43, s4, 23
; SI-NEXT:    v_readfirstlane_b32 s4, v15
; SI-NEXT:    v_writelane_b32 v43, s4, 24
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:80
; SI-NEXT:    v_readfirstlane_b32 s4, v14
; SI-NEXT:    v_writelane_b32 v43, s4, 25
; SI-NEXT:    v_readfirstlane_b32 s4, v13
; SI-NEXT:    v_writelane_b32 v43, s4, 26
; SI-NEXT:    v_readfirstlane_b32 s4, v12
; SI-NEXT:    v_writelane_b32 v43, s4, 27
; SI-NEXT:    v_readfirstlane_b32 s4, v11
; SI-NEXT:    v_writelane_b32 v43, s4, 28
; SI-NEXT:    v_readfirstlane_b32 s4, v10
; SI-NEXT:    v_writelane_b32 v43, s4, 29
; SI-NEXT:    v_readfirstlane_b32 s4, v9
; SI-NEXT:    v_writelane_b32 v43, s4, 30
; SI-NEXT:    v_readfirstlane_b32 s4, v8
; SI-NEXT:    v_writelane_b32 v43, s4, 31
; SI-NEXT:    v_readfirstlane_b32 s4, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:16
; SI-NEXT:    v_writelane_b32 v43, s4, 32
; SI-NEXT:    v_readfirstlane_b32 s4, v1
; SI-NEXT:    v_readfirstlane_b32 s29, v0
; SI-NEXT:    v_readfirstlane_b32 s23, v2
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:48
; SI-NEXT:    v_writelane_b32 v43, s4, 33
; SI-NEXT:    v_writelane_b32 v43, s37, 34
; SI-NEXT:    v_writelane_b32 v43, s73, 35
; SI-NEXT:    v_writelane_b32 v43, s17, 36
; SI-NEXT:    v_writelane_b32 v43, s16, 37
; SI-NEXT:    v_writelane_b32 v43, s21, 38
; SI-NEXT:    v_writelane_b32 v43, s6, 39
; SI-NEXT:    v_writelane_b32 v43, s7, 40
; SI-NEXT:    v_writelane_b32 v43, s75, 41
; SI-NEXT:    v_writelane_b32 v43, s76, 42
; SI-NEXT:    v_writelane_b32 v43, s77, 43
; SI-NEXT:    v_writelane_b32 v43, s78, 44
; SI-NEXT:    v_writelane_b32 v43, s88, 45
; SI-NEXT:    v_writelane_b32 v43, s34, 46
; SI-NEXT:    v_writelane_b32 v43, s64, 47
; SI-NEXT:    v_writelane_b32 v43, s38, 48
; SI-NEXT:    v_writelane_b32 v43, s39, 49
; SI-NEXT:    v_writelane_b32 v43, s48, 50
; SI-NEXT:    v_writelane_b32 v43, s49, 51
; SI-NEXT:    v_writelane_b32 v43, s19, 52
; SI-NEXT:    v_writelane_b32 v43, s95, 53
; SI-NEXT:    v_writelane_b32 v43, s52, 54
; SI-NEXT:    v_writelane_b32 v43, s53, 55
; SI-NEXT:    v_writelane_b32 v43, s54, 56
; SI-NEXT:    v_writelane_b32 v43, s55, 57
; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
; SI-NEXT:    v_writelane_b32 v43, s18, 58
; SI-NEXT:    v_readfirstlane_b32 s90, v3
; SI-NEXT:    v_readfirstlane_b32 s31, v4
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s20, v5
; SI-NEXT:    v_readfirstlane_b32 s26, v6
; SI-NEXT:    v_readfirstlane_b32 s25, v19
; SI-NEXT:    v_readfirstlane_b32 s35, v20
; SI-NEXT:    v_readfirstlane_b32 s15, v21
; SI-NEXT:    v_readfirstlane_b32 s74, v22
; SI-NEXT:    v_readfirstlane_b32 s94, v23
; SI-NEXT:    v_readfirstlane_b32 s46, v24
; SI-NEXT:    v_readfirstlane_b32 s42, v25
; SI-NEXT:    v_readfirstlane_b32 s45, v26
; SI-NEXT:    v_readfirstlane_b32 s57, v27
; SI-NEXT:    v_readfirstlane_b32 s10, v28
; SI-NEXT:    v_readfirstlane_b32 s8, v29
; SI-NEXT:    v_readfirstlane_b32 s12, v30
; SI-NEXT:    v_readfirstlane_b32 s41, v15
; SI-NEXT:    v_readfirstlane_b32 s11, v16
; SI-NEXT:    v_readfirstlane_b32 s27, v17
; SI-NEXT:    v_readfirstlane_b32 s28, v18
; SI-NEXT:    v_readfirstlane_b32 s13, v31
; SI-NEXT:    v_readfirstlane_b32 s99, v32
; SI-NEXT:    v_readfirstlane_b32 s98, v33
; SI-NEXT:    v_readfirstlane_b32 s9, v34
; SI-NEXT:    v_readfirstlane_b32 s85, v7
; SI-NEXT:    v_readfirstlane_b32 s71, v8
; SI-NEXT:    v_readfirstlane_b32 s70, v9
; SI-NEXT:    v_readfirstlane_b32 s80, v10
; SI-NEXT:    v_readfirstlane_b32 s81, v39
; SI-NEXT:    v_readfirstlane_b32 s67, v48
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_readfirstlane_b32 s24, v11
; SI-NEXT:    v_readfirstlane_b32 s68, v50
; SI-NEXT:    v_writelane_b32 v42, s68, 0
; SI-NEXT:    v_readfirstlane_b32 s69, v0
; SI-NEXT:    v_writelane_b32 v42, s69, 1
; SI-NEXT:    v_writelane_b32 v42, s70, 2
; SI-NEXT:    v_writelane_b32 v42, s71, 3
; SI-NEXT:    v_writelane_b32 v42, s80, 4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_readfirstlane_b32 s82, v37
; SI-NEXT:    v_writelane_b32 v42, s81, 5
; SI-NEXT:    v_readfirstlane_b32 s83, v36
; SI-NEXT:    v_writelane_b32 v42, s82, 6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s84, v38
; SI-NEXT:    v_writelane_b32 v42, s83, 7
; SI-NEXT:    v_writelane_b32 v42, s84, 8
; SI-NEXT:    v_readfirstlane_b32 s86, v13
; SI-NEXT:    v_writelane_b32 v42, s85, 9
; SI-NEXT:    v_readfirstlane_b32 s87, v12
; SI-NEXT:    v_writelane_b32 v42, s86, 10
; SI-NEXT:    v_readfirstlane_b32 s96, v14
; SI-NEXT:    v_writelane_b32 v42, s87, 11
; SI-NEXT:    v_readfirstlane_b32 s97, v35
; SI-NEXT:    v_writelane_b32 v42, s96, 12
; SI-NEXT:    v_writelane_b32 v42, s97, 13
; SI-NEXT:    v_writelane_b32 v42, s98, 14
; SI-NEXT:    v_writelane_b32 v42, s99, 15
; SI-NEXT:    v_writelane_b32 v42, s9, 16
; SI-NEXT:    v_writelane_b32 v42, s24, 17
; SI-NEXT:    v_writelane_b32 v42, s27, 18
; SI-NEXT:    v_writelane_b32 v42, s11, 19
; SI-NEXT:    v_writelane_b32 v42, s28, 20
; SI-NEXT:    v_writelane_b32 v42, s13, 21
; SI-NEXT:    v_writelane_b32 v42, s8, 22
; SI-NEXT:    v_writelane_b32 v42, s10, 23
; SI-NEXT:    v_writelane_b32 v42, s12, 24
; SI-NEXT:    v_writelane_b32 v42, s41, 25
; SI-NEXT:    v_writelane_b32 v42, s45, 26
; SI-NEXT:    v_writelane_b32 v42, s57, 27
; SI-NEXT:    v_writelane_b32 v42, s15, 28
; SI-NEXT:    v_writelane_b32 v42, s35, 29
; SI-NEXT:    v_writelane_b32 v42, s74, 30
; SI-NEXT:    v_writelane_b32 v42, s94, 31
; SI-NEXT:    v_writelane_b32 v42, s20, 32
; SI-NEXT:    v_writelane_b32 v42, s31, 33
; SI-NEXT:    v_writelane_b32 v42, s29, 34
; SI-NEXT:    v_writelane_b32 v42, s26, 35
; SI-NEXT:    v_writelane_b32 v42, s25, 36
; SI-NEXT:    v_writelane_b32 v42, s50, 37
; SI-NEXT:    v_writelane_b32 v42, s23, 38
; SI-NEXT:    v_writelane_b32 v42, s36, 39
; SI-NEXT:    v_writelane_b32 v42, s14, 40
; SI-NEXT:    v_writelane_b32 v42, s90, 41
; SI-NEXT:    v_writelane_b32 v42, s72, 42
; SI-NEXT:    v_writelane_b32 v42, s63, 43
; SI-NEXT:    v_writelane_b32 v42, s79, 44
; SI-NEXT:    v_writelane_b32 v42, s89, 45
; SI-NEXT:    v_writelane_b32 v42, s44, 46
; SI-NEXT:    v_writelane_b32 v42, s59, 47
; SI-NEXT:    v_writelane_b32 v42, s60, 48
; SI-NEXT:    v_writelane_b32 v42, s62, 49
; SI-NEXT:    v_writelane_b32 v42, s43, 50
; SI-NEXT:    v_writelane_b32 v42, s91, 51
; SI-NEXT:    v_writelane_b32 v42, s47, 52
; SI-NEXT:    v_writelane_b32 v42, s56, 53
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    v_writelane_b32 v42, s42, 54
; SI-NEXT:    v_readfirstlane_b32 vcc_hi, v1
; SI-NEXT:    v_writelane_b32 v43, vcc_lo, 59
; SI-NEXT:    v_writelane_b32 v42, s46, 55
; SI-NEXT:    v_readfirstlane_b32 s65, v51
; SI-NEXT:    v_writelane_b32 v43, vcc_hi, 60
; SI-NEXT:    v_writelane_b32 v42, s40, 56
; SI-NEXT:    v_readfirstlane_b32 s66, v49
; SI-NEXT:    v_writelane_b32 v43, s65, 61
; SI-NEXT:    v_writelane_b32 v42, s51, 57
; SI-NEXT:    v_writelane_b32 v43, s66, 62
; SI-NEXT:    v_writelane_b32 v42, s30, 58
; SI-NEXT:    v_writelane_b32 v43, s67, 63
; SI-NEXT:    v_writelane_b32 v42, s61, 59
; SI-NEXT:    s_cbranch_scc0 .LBB89_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s5, s17, 24
; SI-NEXT:    s_mov_b32 s16, s21
; SI-NEXT:    s_or_b32 s21, s5, s4
; SI-NEXT:    s_and_b32 s4, s37, 0xff
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s5, s73, 24
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_writelane_b32 v42, s4, 60
; SI-NEXT:    s_and_b32 s4, s6, 0xff
; SI-NEXT:    s_lshl_b32 s5, s16, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_readlane_b32 s5, v43, 7
; SI-NEXT:    s_and_b32 s5, s5, 0xff
; SI-NEXT:    v_readlane_b32 s6, v43, 6
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    v_writelane_b32 v42, s4, 61
; SI-NEXT:    s_or_b32 s4, s6, s5
; SI-NEXT:    v_readlane_b32 s5, v43, 5
; SI-NEXT:    s_and_b32 s5, s5, 0xff
; SI-NEXT:    v_readlane_b32 s6, v43, 4
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    v_writelane_b32 v42, s4, 62
; SI-NEXT:    s_or_b32 s4, s6, s5
; SI-NEXT:    v_readlane_b32 s5, v43, 3
; SI-NEXT:    s_and_b32 s5, s5, 0xff
; SI-NEXT:    v_readlane_b32 s6, v43, 2
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_writelane_b32 v42, s5, 63
; SI-NEXT:    v_readlane_b32 s5, v43, 1
; SI-NEXT:    v_readlane_b32 s6, v43, 0
; SI-NEXT:    s_and_b32 s5, s5, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    s_or_b32 s5, s5, s6
; SI-NEXT:    s_and_b32 s6, s75, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s7, 24
; SI-NEXT:    s_or_b32 s73, s7, s6
; SI-NEXT:    s_and_b32 s6, s52, 0xff
; SI-NEXT:    ; implicit-def: $vgpr41 : SGPR spill to VGPR lane
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s48, 24
; SI-NEXT:    v_writelane_b32 v41, s5, 0
; SI-NEXT:    s_or_b32 s5, s7, s6
; SI-NEXT:    s_and_b32 s6, s77, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s76, 24
; SI-NEXT:    s_or_b32 s76, s7, s6
; SI-NEXT:    s_and_b32 s6, s53, 0xff
; SI-NEXT:    s_lshl_b32 s7, s49, 8
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_and_b32 s7, s88, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_lshl_b32 s16, s78, 24
; SI-NEXT:    s_or_b32 s75, s16, s7
; SI-NEXT:    s_and_b32 s7, s54, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_lshl_b32 s16, s19, 24
; SI-NEXT:    s_or_b32 s49, s16, s7
; SI-NEXT:    s_and_b32 s7, s64, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_lshl_b32 s16, s34, 24
; SI-NEXT:    s_or_b32 s52, s16, s7
; SI-NEXT:    s_and_b32 s7, s55, 0xff
; SI-NEXT:    s_lshl_b32 s16, s95, 8
; SI-NEXT:    s_or_b32 s7, s7, s16
; SI-NEXT:    s_and_b32 s16, s39, 0xff
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s17, s38, 24
; SI-NEXT:    s_or_b32 s54, s17, s16
; SI-NEXT:    s_and_b32 s16, s18, 0xff
; SI-NEXT:    v_readlane_b32 s17, v43, 19
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_or_b32 s38, s17, s16
; SI-NEXT:    v_readlane_b32 s16, v43, 18
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    v_readlane_b32 s17, v43, 17
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_or_b32 s95, s17, s16
; SI-NEXT:    v_readlane_b32 s16, v43, 16
; SI-NEXT:    v_readlane_b32 s17, v43, 15
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v43, 14
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v43, 13
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s53, s18, s17
; SI-NEXT:    v_readlane_b32 s17, v43, 12
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v43, 11
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s55, s18, s17
; SI-NEXT:    v_readlane_b32 s17, v43, 10
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    v_readlane_b32 s18, v43, 9
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s34, s18, s17
; SI-NEXT:    v_readlane_b32 s17, v43, 8
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s18, s65, 8
; SI-NEXT:    s_or_b32 s18, s17, s18
; SI-NEXT:    s_and_b32 s17, vcc_lo, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, vcc_hi, 24
; SI-NEXT:    s_or_b32 s78, s19, s17
; SI-NEXT:    s_and_b32 s17, s69, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s68, 24
; SI-NEXT:    s_or_b32 s64, s19, s17
; SI-NEXT:    s_and_b32 s17, s66, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s67, 24
; SI-NEXT:    s_or_b32 s88, s19, s17
; SI-NEXT:    s_and_b32 s17, s81, 0xff
; SI-NEXT:    s_lshl_b32 s19, s80, 8
; SI-NEXT:    s_or_b32 vcc_lo, s17, s19
; SI-NEXT:    s_and_b32 s17, s70, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s71, 24
; SI-NEXT:    s_mov_b32 s37, s56
; SI-NEXT:    s_or_b32 s56, s19, s17
; SI-NEXT:    s_and_b32 s17, s85, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s84, 24
; SI-NEXT:    s_or_b32 s39, s19, s17
; SI-NEXT:    s_and_b32 s17, s82, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s83, 24
; SI-NEXT:    s_or_b32 s77, s19, s17
; SI-NEXT:    s_and_b32 s17, s97, 0xff
; SI-NEXT:    s_lshl_b32 s19, s96, 8
; SI-NEXT:    s_or_b32 vcc_hi, s17, s19
; SI-NEXT:    s_and_b32 s17, s86, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s87, 24
; SI-NEXT:    s_or_b32 s82, s19, s17
; SI-NEXT:    s_and_b32 s17, s24, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s9, 24
; SI-NEXT:    s_or_b32 s9, s19, s17
; SI-NEXT:    s_and_b32 s17, s98, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s99, 24
; SI-NEXT:    s_or_b32 s48, s19, s17
; SI-NEXT:    s_and_b32 s17, s13, 0xff
; SI-NEXT:    s_lshl_b32 s19, s28, 8
; SI-NEXT:    s_or_b32 s84, s17, s19
; SI-NEXT:    s_and_b32 s17, s27, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s11, 24
; SI-NEXT:    s_or_b32 s13, s19, s17
; SI-NEXT:    s_and_b32 s17, s41, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s12, 24
; SI-NEXT:    s_or_b32 s11, s19, s17
; SI-NEXT:    s_and_b32 s17, s8, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s10, 24
; SI-NEXT:    s_or_b32 s8, s19, s17
; SI-NEXT:    s_and_b32 s17, s57, 0xff
; SI-NEXT:    s_lshl_b32 s19, s45, 8
; SI-NEXT:    s_or_b32 s85, s17, s19
; SI-NEXT:    s_and_b32 s17, s42, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s46, 24
; SI-NEXT:    s_or_b32 s12, s19, s17
; SI-NEXT:    s_and_b32 s17, s94, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s74, 24
; SI-NEXT:    s_or_b32 s10, s19, s17
; SI-NEXT:    s_and_b32 s17, s15, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s35, 24
; SI-NEXT:    s_mov_b32 s46, s21
; SI-NEXT:    s_or_b32 s21, s19, s17
; SI-NEXT:    s_and_b32 s17, s25, 0xff
; SI-NEXT:    s_lshl_b32 s19, s26, 8
; SI-NEXT:    s_or_b32 s86, s17, s19
; SI-NEXT:    s_and_b32 s17, s20, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s31, 24
; SI-NEXT:    s_mov_b32 s42, s49
; SI-NEXT:    s_or_b32 s49, s19, s17
; SI-NEXT:    s_and_b32 s17, s90, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s23, 24
; SI-NEXT:    s_mov_b32 s90, s52
; SI-NEXT:    s_or_b32 s52, s19, s17
; SI-NEXT:    s_and_b32 s17, s29, 0xff
; SI-NEXT:    v_readlane_b32 s15, v43, 33
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s15, 24
; SI-NEXT:    v_readlane_b32 s15, v43, 32
; SI-NEXT:    s_or_b32 s24, s19, s17
; SI-NEXT:    s_and_b32 s17, s15, 0xff
; SI-NEXT:    v_readlane_b32 s15, v43, 31
; SI-NEXT:    s_lshl_b32 s19, s15, 8
; SI-NEXT:    v_readlane_b32 s15, v43, 30
; SI-NEXT:    s_or_b32 s87, s17, s19
; SI-NEXT:    s_and_b32 s17, s15, 0xff
; SI-NEXT:    v_readlane_b32 s15, v43, 29
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s15, 24
; SI-NEXT:    v_readlane_b32 s15, v43, 28
; SI-NEXT:    s_mov_b32 s27, s39
; SI-NEXT:    s_mov_b32 s39, s75
; SI-NEXT:    s_or_b32 s75, s19, s17
; SI-NEXT:    s_and_b32 s17, s15, 0xff
; SI-NEXT:    v_readlane_b32 s15, v43, 27
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s15, 24
; SI-NEXT:    v_readlane_b32 s15, v43, 26
; SI-NEXT:    s_mov_b32 s41, s38
; SI-NEXT:    s_or_b32 s38, s19, s17
; SI-NEXT:    s_and_b32 s17, s15, 0xff
; SI-NEXT:    s_mov_b32 s57, s4
; SI-NEXT:    v_readlane_b32 s4, v43, 25
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s4, 24
; SI-NEXT:    v_readlane_b32 s4, v43, 24
; SI-NEXT:    s_or_b32 s65, s19, s17
; SI-NEXT:    s_and_b32 s17, s4, 0xff
; SI-NEXT:    v_readlane_b32 s4, v43, 23
; SI-NEXT:    s_lshl_b32 s19, s4, 8
; SI-NEXT:    v_readlane_b32 s4, v43, 22
; SI-NEXT:    s_or_b32 s23, s17, s19
; SI-NEXT:    s_and_b32 s17, s4, 0xff
; SI-NEXT:    v_readlane_b32 s4, v43, 21
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s4, 24
; SI-NEXT:    v_readlane_b32 s4, v43, 20
; SI-NEXT:    s_or_b32 s67, s19, s17
; SI-NEXT:    s_and_b32 s17, s4, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s61, 24
; SI-NEXT:    s_or_b32 s66, s19, s17
; SI-NEXT:    s_and_b32 s17, s40, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s58, 24
; SI-NEXT:    s_or_b32 s68, s19, s17
; SI-NEXT:    s_and_b32 s17, s51, 0xff
; SI-NEXT:    s_lshl_b32 s19, s30, 8
; SI-NEXT:    s_or_b32 s25, s17, s19
; SI-NEXT:    s_and_b32 s17, s92, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s93, 24
; SI-NEXT:    s_or_b32 s70, s19, s17
; SI-NEXT:    s_and_b32 s17, s14, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s36, 24
; SI-NEXT:    s_or_b32 s69, s19, s17
; SI-NEXT:    s_and_b32 s17, s50, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s22, 24
; SI-NEXT:    s_or_b32 s71, s19, s17
; SI-NEXT:    s_and_b32 s17, s89, 0xff
; SI-NEXT:    s_lshl_b32 s19, s79, 8
; SI-NEXT:    s_or_b32 s20, s17, s19
; SI-NEXT:    s_and_b32 s17, s72, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s63, 24
; SI-NEXT:    s_or_b32 s81, s19, s17
; SI-NEXT:    s_and_b32 s17, s62, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s60, 24
; SI-NEXT:    s_or_b32 s80, s19, s17
; SI-NEXT:    s_and_b32 s17, s44, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s59, 24
; SI-NEXT:    s_mov_b32 s59, s82
; SI-NEXT:    s_or_b32 s82, s19, s17
; SI-NEXT:    s_and_b32 s17, s37, 0xff
; SI-NEXT:    s_lshl_b32 s19, s47, 8
; SI-NEXT:    s_mov_b32 s94, s4
; SI-NEXT:    s_or_b32 s26, s17, s19
; SI-NEXT:    s_and_b32 s17, s43, 0xff
; SI-NEXT:    v_readlane_b32 s4, v42, 61
; SI-NEXT:    s_mov_b32 s60, s21
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s19, s91, 24
; SI-NEXT:    s_lshl_b32 s21, s4, 16
; SI-NEXT:    v_readlane_b32 s4, v41, 0
; SI-NEXT:    s_mov_b32 s28, s77
; SI-NEXT:    v_readlane_b32 s77, v42, 63
; SI-NEXT:    s_mov_b32 s45, s5
; SI-NEXT:    s_mov_b32 s15, s58
; SI-NEXT:    s_mov_b32 s31, s92
; SI-NEXT:    v_readlane_b32 s74, v42, 62
; SI-NEXT:    s_mov_b32 s29, s93
; SI-NEXT:    v_readlane_b32 s35, v42, 60
; SI-NEXT:    s_mov_b32 s36, s22
; SI-NEXT:    s_or_b32 s83, s19, s17
; SI-NEXT:    s_lshl_b32 s44, s4, 16
; SI-NEXT:    s_lshl_b32 s19, s6, 16
; SI-NEXT:    s_lshl_b32 s17, s7, 16
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_lshl_b32 s99, vcc_lo, 16
; SI-NEXT:    s_lshl_b32 s98, vcc_hi, 16
; SI-NEXT:    s_lshl_b32 s7, s84, 16
; SI-NEXT:    s_lshl_b32 s6, s85, 16
; SI-NEXT:    s_lshl_b32 s97, s86, 16
; SI-NEXT:    s_lshl_b32 s96, s87, 16
; SI-NEXT:    s_lshl_b32 s87, s23, 16
; SI-NEXT:    s_lshl_b32 s86, s25, 16
; SI-NEXT:    s_lshl_b32 s85, s20, 16
; SI-NEXT:    s_lshl_b32 s84, s26, 16
; SI-NEXT:    s_cbranch_execnz .LBB89_3
; SI-NEXT:  .LBB89_2: ; %cmp.true
; SI-NEXT:    v_readlane_b32 s4, v42, 53
; SI-NEXT:    s_add_i32 s59, s4, 3
; SI-NEXT:    v_readlane_b32 s5, v42, 52
; SI-NEXT:    v_readlane_b32 s6, v42, 50
; SI-NEXT:    s_and_b32 s4, s59, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_add_i32 s56, s6, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readlane_b32 s5, v42, 51
; SI-NEXT:    s_and_b32 s6, s56, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 24
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_addk_i32 s4, 0x300
; SI-NEXT:    s_or_b32 s5, s5, s6
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readlane_b32 s5, v42, 49
; SI-NEXT:    s_add_i32 s47, s5, 3
; SI-NEXT:    v_readlane_b32 s6, v42, 48
; SI-NEXT:    v_readlane_b32 s7, v42, 46
; SI-NEXT:    s_and_b32 s5, s47, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    s_add_i32 s44, s7, 3
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_readlane_b32 s6, v42, 47
; SI-NEXT:    s_and_b32 s7, s44, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_addk_i32 s5, 0x300
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_readlane_b32 s6, v42, 45
; SI-NEXT:    s_add_i32 s43, s6, 3
; SI-NEXT:    v_readlane_b32 s7, v42, 44
; SI-NEXT:    v_readlane_b32 s8, v42, 42
; SI-NEXT:    s_and_b32 s6, s43, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_add_i32 s40, s8, 3
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readlane_b32 s7, v42, 43
; SI-NEXT:    s_and_b32 s16, s40, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_addk_i32 s6, 0x300
; SI-NEXT:    s_or_b32 s7, s7, s16
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readlane_b32 s7, v42, 40
; SI-NEXT:    v_readlane_b32 s8, v42, 39
; SI-NEXT:    s_add_i32 s14, s7, 3
; SI-NEXT:    s_lshl_b32 s13, s8, 8
; SI-NEXT:    v_readlane_b32 s8, v42, 37
; SI-NEXT:    s_and_b32 s7, s14, 0xff
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    s_or_b32 s7, s13, s7
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s9, s36, 24
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_addk_i32 s7, 0x300
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_or_b32 s9, s8, s7
; SI-NEXT:    v_readlane_b32 s7, v42, 57
; SI-NEXT:    s_add_i32 s24, s7, 3
; SI-NEXT:    v_readlane_b32 s8, v42, 58
; SI-NEXT:    s_and_b32 s7, s24, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_add_i32 s13, s31, 3
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s8, s29, 24
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_addk_i32 s7, 0x300
; SI-NEXT:    s_or_b32 s8, s8, s13
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    v_readlane_b32 s10, v42, 59
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    s_add_i32 s8, s94, 3
; SI-NEXT:    s_lshl_b32 s13, s10, 8
; SI-NEXT:    v_readlane_b32 s10, v42, 56
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_add_i32 s14, s10, 3
; SI-NEXT:    s_or_b32 s8, s13, s8
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s13, s15, 24
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s13, s13, s14
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s13, s13, s8
; SI-NEXT:    v_readlane_b32 s8, v43, 24
; SI-NEXT:    s_add_i32 s27, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v43, 23
; SI-NEXT:    s_and_b32 s8, s27, 0xff
; SI-NEXT:    s_lshl_b32 s14, s10, 8
; SI-NEXT:    v_readlane_b32 s10, v43, 21
; SI-NEXT:    s_or_b32 s8, s14, s8
; SI-NEXT:    s_lshl_b32 s14, s10, 24
; SI-NEXT:    v_readlane_b32 s10, v43, 22
; SI-NEXT:    s_add_i32 s16, s10, 3
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s14, s14, s8
; SI-NEXT:    v_readlane_b32 s8, v43, 28
; SI-NEXT:    s_add_i32 s10, s8, 3
; SI-NEXT:    s_and_b32 s8, s10, 0xff
; SI-NEXT:    v_readlane_b32 s10, v43, 27
; SI-NEXT:    v_readlane_b32 s11, v43, 26
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s16, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v43, 25
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s16
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s16, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v43, 32
; SI-NEXT:    s_add_i32 s12, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v43, 31
; SI-NEXT:    v_readlane_b32 s11, v43, 30
; SI-NEXT:    s_and_b32 s8, s12, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s12, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v43, 29
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s17, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v42, 41
; SI-NEXT:    s_add_i32 s15, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v42, 38
; SI-NEXT:    v_readlane_b32 s11, v42, 34
; SI-NEXT:    s_and_b32 s8, s15, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s20, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v43, 33
; SI-NEXT:    s_and_b32 s11, s20, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s15, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v42, 36
; SI-NEXT:    s_add_i32 s28, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v42, 35
; SI-NEXT:    v_readlane_b32 s11, v42, 32
; SI-NEXT:    s_and_b32 s8, s28, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v42, 33
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s20, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v42, 31
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v42, 30
; SI-NEXT:    v_readlane_b32 s11, v42, 28
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v42, 29
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s23, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v42, 27
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v42, 26
; SI-NEXT:    v_readlane_b32 s11, v42, 54
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v42, 55
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s18, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v42, 25
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v42, 24
; SI-NEXT:    v_readlane_b32 s11, v42, 22
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v42, 23
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s24, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v42, 21
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    v_readlane_b32 s10, v42, 20
; SI-NEXT:    v_readlane_b32 s11, v42, 18
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v42, 19
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_add_i32 s11, s13, 0x3000000
; SI-NEXT:    s_add_i32 s13, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v42, 17
; SI-NEXT:    s_or_b32 s19, s10, s8
; SI-NEXT:    s_add_i32 s8, s6, 0x3000000
; SI-NEXT:    s_add_i32 s12, s14, 0x3000000
; SI-NEXT:    s_add_i32 s14, s17, 0x3000000
; SI-NEXT:    s_add_i32 s6, s18, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v42, 16
; SI-NEXT:    v_readlane_b32 s18, v42, 14
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 15
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s25, s17, s16
; SI-NEXT:    v_readlane_b32 s16, v42, 13
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v42, 12
; SI-NEXT:    v_readlane_b32 s18, v42, 10
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 11
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s26, s17, s16
; SI-NEXT:    v_readlane_b32 s16, v42, 9
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v42, 8
; SI-NEXT:    v_readlane_b32 s18, v42, 6
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 7
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s27, s17, s16
; SI-NEXT:    v_readlane_b32 s16, v42, 5
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v42, 4
; SI-NEXT:    v_readlane_b32 s18, v42, 2
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 3
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s28, s17, s16
; SI-NEXT:    v_readlane_b32 s16, v42, 1
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v42, 0
; SI-NEXT:    v_readlane_b32 s18, v43, 62
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 63
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s29, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 8
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 61
; SI-NEXT:    v_readlane_b32 s18, v43, 59
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 60
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s18, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 12
; SI-NEXT:    s_add_i32 s10, s7, 0x3000000
; SI-NEXT:    s_add_i32 s7, s19, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 11
; SI-NEXT:    v_readlane_b32 s19, v43, 10
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 9
; SI-NEXT:    s_and_b32 s19, s19, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s19
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s40, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 16
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 15
; SI-NEXT:    v_readlane_b32 s19, v43, 14
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 13
; SI-NEXT:    s_and_b32 s19, s19, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s19
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 58
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    v_readlane_b32 s19, v43, 19
; SI-NEXT:    v_readlane_b32 s21, v43, 18
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s19, s19, 8
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_or_b32 s17, s19, s17
; SI-NEXT:    v_readlane_b32 s19, v43, 17
; SI-NEXT:    s_and_b32 s21, s21, 0xff
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_lshl_b32 s19, s19, 24
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s19, s19, s21
; SI-NEXT:    s_or_b32 s17, s19, s17
; SI-NEXT:    s_add_i32 s41, s17, 0x3000000
; SI-NEXT:    v_readlane_b32 s17, v43, 57
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    v_readlane_b32 s19, v43, 53
; SI-NEXT:    v_readlane_b32 s21, v43, 49
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s19, s19, 8
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_or_b32 s17, s19, s17
; SI-NEXT:    v_readlane_b32 s19, v43, 48
; SI-NEXT:    s_and_b32 s21, s21, 0xff
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_lshl_b32 s19, s19, 24
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s19, s19, s21
; SI-NEXT:    s_or_b32 s17, s19, s17
; SI-NEXT:    v_readlane_b32 s19, v43, 56
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    v_readlane_b32 s21, v43, 52
; SI-NEXT:    v_readlane_b32 s22, v43, 47
; SI-NEXT:    s_and_b32 s19, s19, 0xff
; SI-NEXT:    s_lshl_b32 s21, s21, 8
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_or_b32 s19, s21, s19
; SI-NEXT:    v_readlane_b32 s21, v43, 46
; SI-NEXT:    s_and_b32 s22, s22, 0xff
; SI-NEXT:    s_addk_i32 s19, 0x300
; SI-NEXT:    s_lshl_b32 s21, s21, 24
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_or_b32 s19, s21, s19
; SI-NEXT:    s_add_i32 s42, s19, 0x3000000
; SI-NEXT:    v_readlane_b32 s19, v43, 55
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    v_readlane_b32 s21, v43, 51
; SI-NEXT:    v_readlane_b32 s22, v43, 45
; SI-NEXT:    s_and_b32 s19, s19, 0xff
; SI-NEXT:    s_lshl_b32 s21, s21, 8
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_or_b32 s19, s21, s19
; SI-NEXT:    v_readlane_b32 s21, v43, 44
; SI-NEXT:    s_and_b32 s22, s22, 0xff
; SI-NEXT:    s_addk_i32 s19, 0x300
; SI-NEXT:    s_lshl_b32 s21, s21, 24
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_or_b32 s19, s21, s19
; SI-NEXT:    v_readlane_b32 s21, v43, 54
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    v_readlane_b32 s22, v43, 50
; SI-NEXT:    v_readlane_b32 s43, v43, 43
; SI-NEXT:    s_and_b32 s21, s21, 0xff
; SI-NEXT:    s_lshl_b32 s22, s22, 8
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    v_readlane_b32 s22, v43, 42
; SI-NEXT:    s_and_b32 s43, s43, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s22, s22, 24
; SI-NEXT:    s_lshl_b32 s43, s43, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s43
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    s_add_i32 s43, s21, 0x3000000
; SI-NEXT:    v_readlane_b32 s21, v43, 1
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    v_readlane_b32 s22, v43, 0
; SI-NEXT:    v_readlane_b32 s44, v43, 41
; SI-NEXT:    s_and_b32 s21, s21, 0xff
; SI-NEXT:    s_lshl_b32 s22, s22, 8
; SI-NEXT:    s_add_i32 s44, s44, 3
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    v_readlane_b32 s22, v43, 40
; SI-NEXT:    s_and_b32 s44, s44, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s22, s22, 24
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s22, s22, s44
; SI-NEXT:    s_or_b32 s21, s22, s21
; SI-NEXT:    s_add_i32 s22, s21, 0x3000000
; SI-NEXT:    v_readlane_b32 s21, v43, 5
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    v_readlane_b32 s44, v43, 4
; SI-NEXT:    v_readlane_b32 s45, v43, 3
; SI-NEXT:    s_and_b32 s21, s21, 0xff
; SI-NEXT:    s_lshl_b32 s44, s44, 8
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_or_b32 s21, s44, s21
; SI-NEXT:    v_readlane_b32 s44, v43, 2
; SI-NEXT:    s_and_b32 s45, s45, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s44, s44, 24
; SI-NEXT:    s_lshl_b32 s45, s45, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s44, s44, s45
; SI-NEXT:    s_or_b32 s21, s44, s21
; SI-NEXT:    s_add_i32 s44, s21, 0x3000000
; SI-NEXT:    v_readlane_b32 s21, v43, 39
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    v_readlane_b32 s45, v43, 38
; SI-NEXT:    v_readlane_b32 s46, v43, 7
; SI-NEXT:    s_and_b32 s21, s21, 0xff
; SI-NEXT:    s_lshl_b32 s45, s45, 8
; SI-NEXT:    s_add_i32 s46, s46, 3
; SI-NEXT:    s_or_b32 s21, s45, s21
; SI-NEXT:    v_readlane_b32 s45, v43, 6
; SI-NEXT:    s_and_b32 s46, s46, 0xff
; SI-NEXT:    s_addk_i32 s21, 0x300
; SI-NEXT:    s_lshl_b32 s45, s45, 24
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_or_b32 s45, s45, s46
; SI-NEXT:    s_or_b32 s21, s45, s21
; SI-NEXT:    v_readlane_b32 s45, v43, 37
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    v_readlane_b32 s46, v43, 36
; SI-NEXT:    v_readlane_b32 s47, v43, 34
; SI-NEXT:    s_and_b32 s45, s45, 0xff
; SI-NEXT:    s_lshl_b32 s46, s46, 8
; SI-NEXT:    s_add_i32 s47, s47, 3
; SI-NEXT:    s_or_b32 s45, s46, s45
; SI-NEXT:    v_readlane_b32 s46, v43, 35
; SI-NEXT:    s_and_b32 s47, s47, 0xff
; SI-NEXT:    s_addk_i32 s45, 0x300
; SI-NEXT:    s_lshl_b32 s46, s46, 24
; SI-NEXT:    s_lshl_b32 s47, s47, 16
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_or_b32 s46, s46, s47
; SI-NEXT:    s_add_i32 s23, s23, 0x3000000
; SI-NEXT:    s_add_i32 s24, s24, 0x3000000
; SI-NEXT:    s_add_i32 s25, s25, 0x3000000
; SI-NEXT:    s_or_b32 s45, s46, s45
; SI-NEXT:    s_add_i32 s4, s4, 0x3000000
; SI-NEXT:    s_add_i32 s5, s5, 0x3000000
; SI-NEXT:    s_add_i32 s9, s9, 0x3000000
; SI-NEXT:    s_add_i32 s15, s15, 0x3000000
; SI-NEXT:    s_add_i32 s20, s20, 0x3000000
; SI-NEXT:    s_add_i32 s26, s26, 0x3000000
; SI-NEXT:    s_add_i32 s27, s27, 0x3000000
; SI-NEXT:    s_add_i32 s28, s28, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 0x3000000
; SI-NEXT:    s_add_i32 s17, s17, 0x3000000
; SI-NEXT:    s_add_i32 s19, s19, 0x3000000
; SI-NEXT:    s_add_i32 s21, s21, 0x3000000
; SI-NEXT:    s_add_i32 s45, s45, 0x3000000
; SI-NEXT:    s_and_b32 s77, s44, 0xffff0000
; SI-NEXT:    s_lshl_b32 s57, s44, 16
; SI-NEXT:    s_and_b32 s73, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s44, s22, 16
; SI-NEXT:    s_and_b32 s34, s40, 0xffff0000
; SI-NEXT:    s_lshl_b32 s55, s40, 16
; SI-NEXT:    s_and_b32 s88, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s64, s29, 16
; SI-NEXT:    s_and_b32 s48, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s25, 16
; SI-NEXT:    s_and_b32 s29, s7, 0xffff0000
; SI-NEXT:    s_and_b32 s58, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s25, s24, 16
; SI-NEXT:    s_and_b32 s40, s6, 0xffff0000
; SI-NEXT:    s_and_b32 s60, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s35, s45, 0xffff0000
; SI-NEXT:    s_lshl_b32 s46, s45, 16
; SI-NEXT:    s_and_b32 s74, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s76, s43, 0xffff0000
; SI-NEXT:    s_lshl_b32 s45, s43, 16
; SI-NEXT:    s_and_b32 s39, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s90, s42, 0xffff0000
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_and_b32 s54, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_and_b32 s95, s41, 0xffff0000
; SI-NEXT:    s_lshl_b32 s41, s41, 16
; SI-NEXT:    s_and_b32 s53, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_and_b32 s78, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s56, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s99, s28, 16
; SI-NEXT:    s_and_b32 s28, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s27, s27, 16
; SI-NEXT:    s_and_b32 s59, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s98, s26, 16
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_and_b32 s49, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s97, s20, 16
; SI-NEXT:    s_and_b32 s24, s15, 0xffff0000
; SI-NEXT:    s_lshl_b32 s52, s15, 16
; SI-NEXT:    s_and_b32 s75, s14, 0xffff0000
; SI-NEXT:    s_lshl_b32 s96, s14, 16
; SI-NEXT:    s_and_b32 s65, s13, 0xffff0000
; SI-NEXT:    s_lshl_b32 s38, s13, 16
; SI-NEXT:    s_mov_b32 s13, s29
; SI-NEXT:    s_and_b32 s67, s12, 0xffff0000
; SI-NEXT:    s_lshl_b32 s87, s12, 16
; SI-NEXT:    s_mov_b32 s12, s40
; SI-NEXT:    s_and_b32 s68, s11, 0xffff0000
; SI-NEXT:    s_lshl_b32 s66, s11, 16
; SI-NEXT:    s_mov_b32 s11, s25
; SI-NEXT:    s_and_b32 s70, s10, 0xffff0000
; SI-NEXT:    s_lshl_b32 s86, s10, 16
; SI-NEXT:    s_mov_b32 s10, s23
; SI-NEXT:    s_and_b32 s71, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s69, s9, 16
; SI-NEXT:    s_mov_b32 s9, s22
; SI-NEXT:    s_and_b32 s81, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s85, s8, 16
; SI-NEXT:    s_mov_b32 s8, s58
; SI-NEXT:    s_and_b32 s82, s5, 0xffff0000
; SI-NEXT:    s_lshl_b32 s80, s5, 16
; SI-NEXT:    s_and_b32 s83, s4, 0xffff0000
; SI-NEXT:    s_lshl_b32 s84, s4, 16
; SI-NEXT:  .LBB89_3: ; %end
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s35
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s46
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s74
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s21
; SI-NEXT:    v_lshr_b64 v[1:2], v[1:2], 16
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s77
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s57
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s73
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s44
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s76
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s45
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s39
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s19
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s90
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s42
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s54
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s17
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s95
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s41
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s53
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s16
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s34
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s55
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s78
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s18
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s88
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s64
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s56
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s99
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s28
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s27
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s59
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s98
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s48
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s9
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s13
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s7
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s8
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s11
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s12
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s60
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s10
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s49
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s97
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s24
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s52
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s75
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s96
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s65
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s38
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s67
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s87
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s68
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v26
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s66
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s70
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v27
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s86
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s71
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v28
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s69
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s81
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v29
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s85
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s82
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v30
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s80
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s83
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v31
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s84
; SI-NEXT:    v_lshr_b64 v[31:32], v[31:32], 16
; SI-NEXT:    v_readlane_b32 s99, v40, 35
; SI-NEXT:    v_readlane_b32 s98, v40, 34
; SI-NEXT:    v_readlane_b32 s97, v40, 33
; SI-NEXT:    v_readlane_b32 s96, v40, 32
; SI-NEXT:    v_readlane_b32 s87, v40, 31
; SI-NEXT:    v_readlane_b32 s86, v40, 30
; SI-NEXT:    v_readlane_b32 s85, v40, 29
; SI-NEXT:    v_readlane_b32 s84, v40, 28
; SI-NEXT:    v_readlane_b32 s83, v40, 27
; SI-NEXT:    v_readlane_b32 s82, v40, 26
; SI-NEXT:    v_readlane_b32 s81, v40, 25
; SI-NEXT:    v_readlane_b32 s80, v40, 24
; SI-NEXT:    v_readlane_b32 s71, v40, 23
; SI-NEXT:    v_readlane_b32 s70, v40, 22
; SI-NEXT:    v_readlane_b32 s69, v40, 21
; SI-NEXT:    v_readlane_b32 s68, v40, 20
; SI-NEXT:    v_readlane_b32 s67, v40, 19
; SI-NEXT:    v_readlane_b32 s66, v40, 18
; SI-NEXT:    v_readlane_b32 s65, v40, 17
; SI-NEXT:    v_readlane_b32 s64, v40, 16
; SI-NEXT:    v_readlane_b32 s55, v40, 15
; SI-NEXT:    v_readlane_b32 s54, v40, 14
; SI-NEXT:    v_readlane_b32 s53, v40, 13
; SI-NEXT:    v_readlane_b32 s52, v40, 12
; SI-NEXT:    v_readlane_b32 s51, v40, 11
; SI-NEXT:    v_readlane_b32 s50, v40, 10
; SI-NEXT:    v_readlane_b32 s49, v40, 9
; SI-NEXT:    v_readlane_b32 s48, v40, 8
; SI-NEXT:    v_readlane_b32 s39, v40, 7
; SI-NEXT:    v_readlane_b32 s38, v40, 6
; SI-NEXT:    v_readlane_b32 s37, v40, 5
; SI-NEXT:    v_readlane_b32 s36, v40, 4
; SI-NEXT:    v_readlane_b32 s35, v40, 3
; SI-NEXT:    v_readlane_b32 s34, v40, 2
; SI-NEXT:    v_readlane_b32 s31, v40, 1
; SI-NEXT:    v_readlane_b32 s30, v40, 0
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB89_4:
; SI-NEXT:    s_mov_b32 s36, s22
; SI-NEXT:    s_mov_b32 s29, s93
; SI-NEXT:    s_mov_b32 s31, s92
; SI-NEXT:    v_readlane_b32 s94, v43, 20
; SI-NEXT:    s_mov_b32 s15, s58
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr35
; SI-NEXT:    ; implicit-def: $sgpr21
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr57
; SI-NEXT:    ; implicit-def: $sgpr77
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr73
; SI-NEXT:    ; implicit-def: $sgpr45
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr19
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr17
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr41
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr16
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr18
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    ; implicit-def: $sgpr99
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr27
; SI-NEXT:    ; implicit-def: $sgpr28
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr59
; SI-NEXT:    ; implicit-def: $sgpr9
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr7
; SI-NEXT:    ; implicit-def: $sgpr13
; SI-NEXT:    ; implicit-def: $sgpr11
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr97
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr24
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr75
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr87
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr71
; SI-NEXT:    ; implicit-def: $sgpr85
; SI-NEXT:    ; implicit-def: $sgpr81
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr83
; SI-NEXT:    s_branch .LBB89_2
;
; VI-LABEL: bitcast_v128i8_to_v64bf16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s63, v12
; VI-NEXT:    v_readfirstlane_b32 s8, v0
; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:280
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:12
; VI-NEXT:    v_readfirstlane_b32 s59, v29
; VI-NEXT:    v_readfirstlane_b32 s47, v27
; VI-NEXT:    v_readfirstlane_b32 s75, v30
; VI-NEXT:    v_readfirstlane_b32 s72, v28
; VI-NEXT:    v_readfirstlane_b32 s60, v26
; VI-NEXT:    v_readfirstlane_b32 s43, v25
; VI-NEXT:    v_readfirstlane_b32 s56, v24
; VI-NEXT:    v_readfirstlane_b32 s15, v23
; VI-NEXT:    v_readfirstlane_b32 s44, v22
; VI-NEXT:    v_readfirstlane_b32 s12, v21
; VI-NEXT:    v_readfirstlane_b32 s40, v20
; VI-NEXT:    v_readfirstlane_b32 s76, v18
; VI-NEXT:    v_readfirstlane_b32 s73, v16
; VI-NEXT:    v_readfirstlane_b32 s9, v19
; VI-NEXT:    v_readfirstlane_b32 s61, v17
; VI-NEXT:    v_readfirstlane_b32 s62, v15
; VI-NEXT:    v_readfirstlane_b32 s74, v14
; VI-NEXT:    v_readfirstlane_b32 s57, v13
; VI-NEXT:    v_readfirstlane_b32 s58, v10
; VI-NEXT:    v_readfirstlane_b32 s41, v9
; VI-NEXT:    v_readfirstlane_b32 s46, v8
; VI-NEXT:    v_readfirstlane_b32 s13, v7
; VI-NEXT:    v_readfirstlane_b32 s42, v6
; VI-NEXT:    v_readfirstlane_b32 s10, v5
; VI-NEXT:    v_readfirstlane_b32 s14, v4
; VI-NEXT:    v_readfirstlane_b32 s7, v3
; VI-NEXT:    v_readfirstlane_b32 s11, v2
; VI-NEXT:    v_readfirstlane_b32 s6, v1
; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:328
; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:316
; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:312
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:308
; VI-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:304
; VI-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:300
; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:292
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:288
; VI-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:272
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:268
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:252
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    s_and_b64 s[4:5], vcc, exec
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:216
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:212
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:208
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:200
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:192
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:188
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:184
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:180
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:176
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:172
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:168
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:164
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:160
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:156
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:152
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:148
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:144
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:140
; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:136
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:132
; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:112
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:64
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:40
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    s_cbranch_scc0 .LBB89_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_mov_b32_e32 v3, s23
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v2, s21
; VI-NEXT:    v_perm_b32 v3, s22, v3, v11
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_perm_b32 v2, s20, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_mov_b32_e32 v4, s6
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    v_perm_b32 v4, s8, v4, v11
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_mov_b32_e32 v5, s10
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_mov_b32_e32 v4, s7
; VI-NEXT:    v_perm_b32 v5, s14, v5, v11
; VI-NEXT:    v_perm_b32 v4, s11, v4, v11
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_mov_b32_e32 v6, s41
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_mov_b32_e32 v5, s13
; VI-NEXT:    v_perm_b32 v6, s46, v6, v11
; VI-NEXT:    v_perm_b32 v5, s42, v5, v11
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_mov_b32_e32 v7, s57
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_mov_b32_e32 v6, s45
; VI-NEXT:    v_perm_b32 v7, s63, v7, v11
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, s61
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    v_mov_b32_e32 v7, s62
; VI-NEXT:    v_perm_b32 v8, s73, v8, v11
; VI-NEXT:    v_perm_b32 v7, s74, v7, v11
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v9, s12
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_mov_b32_e32 v8, s9
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_perm_b32 v8, s76, v8, v11
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_mov_b32_e32 v43, v10
; VI-NEXT:    v_mov_b32_e32 v10, s43
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_perm_b32 v10, s56, v10, v11
; VI-NEXT:    v_perm_b32 v9, s44, v9, v11
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_mov_b32_e32 v35, v12
; VI-NEXT:    v_mov_b32_e32 v12, s59
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_mov_b32_e32 v10, s47
; VI-NEXT:    v_perm_b32 v12, s72, v12, v11
; VI-NEXT:    v_perm_b32 v10, s60, v10, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_or_b32_e32 v10, v10, v12
; VI-NEXT:    v_perm_b32 v12, v15, v14, s4
; VI-NEXT:    v_perm_b32 v11, s75, v13, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v18, v23, v18, s4
; VI-NEXT:    v_perm_b32 v23, v51, v33, s4
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_perm_b32 v12, v56, v12, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v13, v47, v13, s4
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v14, v59, v14, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v15, v32, v15, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v61, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_perm_b32 v14, v17, v60, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_perm_b32 v15, v21, v16, s4
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v17, v28, v24, s4
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v21, v58, v46, s4
; VI-NEXT:    v_perm_b32 v24, v36, v49, s4
; VI-NEXT:    v_mov_b32_e32 v28, v52
; VI-NEXT:    v_mov_b32_e32 v52, v53
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v19, v16, s4
; VI-NEXT:    v_perm_b32 v19, v26, v22, s4
; VI-NEXT:    v_or_b32_e32 v17, v19, v17
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    v_perm_b32 v16, v25, v20, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_or_b32_e32 v16, v18, v16
; VI-NEXT:    v_perm_b32 v18, v45, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v20, v44, v29, s4
; VI-NEXT:    v_perm_b32 v22, v53, v42, s4
; VI-NEXT:    v_mov_b32_e32 v45, v34
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_perm_b32 v19, v30, v19, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_perm_b32 v19, v55, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_perm_b32 v20, v39, v57, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v20, v21, v20
; VI-NEXT:    v_perm_b32 v21, v38, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v21, v22, v21
; VI-NEXT:    v_perm_b32 v22, v54, v37, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v22, v23, v22
; VI-NEXT:    v_perm_b32 v23, v40, v50, s4
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v23, v24, v23
; VI-NEXT:    v_perm_b32 v24, v48, v34, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_mov_b32_e32 v48, v25
; VI-NEXT:    v_perm_b32 v25, v25, v41, s4
; VI-NEXT:    v_or_b32_e32 v24, v25, v24
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v34, v33
; VI-NEXT:    v_mov_b32_e32 v33, v26
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v55, v57
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_mov_b32_e32 v53, v25
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    v_or_b32_e32 v25, v26, v25
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v29, v27, s4
; VI-NEXT:    v_or_b32_e32 v26, v27, v26
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v29, v27, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    v_or_b32_e32 v27, v28, v27
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v28, v28, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v62, v29, s4
; VI-NEXT:    v_or_b32_e32 v28, v29, v28
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, v44
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v43, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v29, v30, v29
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v35, v31, s4
; VI-NEXT:    v_or_b32_e32 v30, v31, v30
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v35, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_mov_b32_e32 v35, v43
; VI-NEXT:    v_perm_b32 v43, v44, v43, s4
; VI-NEXT:    v_or_b32_e32 v31, v43, v31
; VI-NEXT:    v_mov_b32_e32 v43, v39
; VI-NEXT:    v_mov_b32_e32 v39, v34
; VI-NEXT:    v_mov_b32_e32 v34, v52
; VI-NEXT:    v_mov_b32_e32 v52, v45
; VI-NEXT:    s_cbranch_execnz .LBB89_3
; VI-NEXT:  .LBB89_2: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_add_i32 s75, s75, 3
; VI-NEXT:    s_add_i32 s72, s72, 3
; VI-NEXT:    s_add_i32 s60, s60, 3
; VI-NEXT:    s_add_i32 s56, s56, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s76, s76, 3
; VI-NEXT:    s_add_i32 s73, s73, 3
; VI-NEXT:    s_add_i32 s74, s74, 3
; VI-NEXT:    s_add_i32 s63, s63, 3
; VI-NEXT:    s_add_i32 s58, s58, 3
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    v_mov_b32_e32 v18, s17
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v62
; VI-NEXT:    v_perm_b32 v0, v0, v35, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, s21
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v35, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v0, v1, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v33
; VI-NEXT:    v_perm_b32 v25, v0, v53, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x300, v28
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_sdwa v26, v26, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x3000000, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x300, v29
; VI-NEXT:    v_or_b32_sdwa v25, v25, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x3000000, v25
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_perm_b32 v24, v0, v52, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v48
; VI-NEXT:    v_perm_b32 v30, v0, v41, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v40
; VI-NEXT:    v_perm_b32 v23, v0, v50, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v36
; VI-NEXT:    v_perm_b32 v31, v0, v49, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v54
; VI-NEXT:    v_perm_b32 v22, v0, v37, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v51
; VI-NEXT:    v_perm_b32 v33, v0, v39, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v38
; VI-NEXT:    v_perm_b32 v21, v0, v63, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v34
; VI-NEXT:    v_perm_b32 v36, v0, v42, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v43
; VI-NEXT:    v_perm_b32 v20, v0, v55, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v58
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v63, s19
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x300, v30
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x300, v31
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x300, v33
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v23, v23, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v24, v24, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x3000000, v22
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x3000000, v23
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x3000000, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v52, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v53, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v37, vcc, 0x300, v53
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v38, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v34, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v10, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v14, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x300, v15
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_perm_b32 v16, v0, v60, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v61
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x300, v34
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v54, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v59
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v40, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v56
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v42, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v47
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v43, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v44, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v0, 0xc0c0004
; VI-NEXT:    v_perm_b32 v62, s20, v62, v0
; VI-NEXT:    v_perm_b32 v63, s18, v63, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v45, s75, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s59
; VI-NEXT:    v_perm_b32 v55, s72, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_perm_b32 v41, s60, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s43
; VI-NEXT:    v_perm_b32 v9, s56, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s15
; VI-NEXT:    v_perm_b32 v46, s44, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s12
; VI-NEXT:    v_perm_b32 v8, s40, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s9
; VI-NEXT:    v_perm_b32 v47, s76, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s61
; VI-NEXT:    v_perm_b32 v7, s73, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s62
; VI-NEXT:    v_perm_b32 v56, s74, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s57
; VI-NEXT:    v_perm_b32 v6, s63, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s45
; VI-NEXT:    v_perm_b32 v57, s58, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s41
; VI-NEXT:    v_perm_b32 v5, s46, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s13
; VI-NEXT:    v_perm_b32 v58, s42, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s10
; VI-NEXT:    v_perm_b32 v4, s14, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s7
; VI-NEXT:    v_perm_b32 v59, s11, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_perm_b32 v3, s8, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s29
; VI-NEXT:    v_perm_b32 v60, s28, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s27
; VI-NEXT:    v_perm_b32 v2, s26, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s25
; VI-NEXT:    v_perm_b32 v61, s24, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s23
; VI-NEXT:    v_perm_b32 v1, s22, v1, v0
; VI-NEXT:    v_perm_b32 v0, s16, v18, v0
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v63
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
; VI-NEXT:    v_or_b32_sdwa v0, v18, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v62
; VI-NEXT:    v_or_b32_sdwa v1, v1, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v61
; VI-NEXT:    v_or_b32_sdwa v2, v2, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v60
; VI-NEXT:    v_or_b32_sdwa v3, v3, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v59
; VI-NEXT:    v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v58
; VI-NEXT:    v_or_b32_sdwa v5, v5, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v57
; VI-NEXT:    v_or_b32_sdwa v6, v6, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v56
; VI-NEXT:    v_or_b32_sdwa v7, v7, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v47
; VI-NEXT:    v_or_b32_sdwa v8, v8, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v46
; VI-NEXT:    v_or_b32_sdwa v9, v9, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
; VI-NEXT:    v_add_u32_e32 v46, vcc, 0x300, v11
; VI-NEXT:    v_add_u32_e32 v56, vcc, 0x300, v13
; VI-NEXT:    v_lshlrev_b32_e32 v57, 16, v14
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v16
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x300, v17
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v54
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x300, v40
; VI-NEXT:    v_or_b32_sdwa v13, v13, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v17, v18, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v47, 16, v12
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v42
; VI-NEXT:    v_add_u32_e32 v54, vcc, 0x300, v43
; VI-NEXT:    v_or_b32_sdwa v12, v12, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v55
; VI-NEXT:    v_add_u32_e32 v55, vcc, 0x300, v41
; VI-NEXT:    v_or_b32_sdwa v10, v10, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v44
; VI-NEXT:    v_add_u32_e32 v40, vcc, 0x300, v45
; VI-NEXT:    v_or_b32_sdwa v11, v11, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v43, vcc, 0x300, v35
; VI-NEXT:    v_add_u32_e32 v35, vcc, 0x300, v36
; VI-NEXT:    v_add_u32_e32 v36, vcc, 0x300, v19
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v52
; VI-NEXT:    v_or_b32_sdwa v14, v14, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v15, v57, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v16, v47, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v19, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v20, v20, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v21, v21, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x3000000, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x3000000, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x3000000, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v6
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v7
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x3000000, v8
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x3000000, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v10
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x3000000, v11
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x3000000, v12
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x3000000, v13
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x3000000, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x3000000, v15
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x3000000, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x3000000, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x3000000, v19
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v20
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x3000000, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v54, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v55, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v31, v54, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x3000000, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v40, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v41, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v30, v40, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x3000000, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v42, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v29, v42, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x3000000, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v44, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v45, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v28, v44, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x3000000, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v46, 16, v18
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v38
; VI-NEXT:    v_or_b32_sdwa v18, v18, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v27, v46, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x3000000, v18
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x3000000, v27
; VI-NEXT:  .LBB89_3: ; %end
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB89_4:
; VI-NEXT:    v_mov_b32_e32 v43, v39
; VI-NEXT:    v_mov_b32_e32 v52, v34
; VI-NEXT:    v_mov_b32_e32 v39, v33
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v34, v53
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, v2
; VI-NEXT:    v_mov_b32_e32 v35, v1
; VI-NEXT:    v_mov_b32_e32 v55, v57
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_branch .LBB89_2
;
; GFX9-LABEL: bitcast_v128i8_to_v64bf16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s63, v10
; GFX9-NEXT:    v_readfirstlane_b32 s7, v0
; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    v_readfirstlane_b32 s43, v2
; GFX9-NEXT:    v_readfirstlane_b32 s6, v1
; GFX9-NEXT:    v_readfirstlane_b32 s8, v5
; GFX9-NEXT:    v_readfirstlane_b32 s10, v4
; GFX9-NEXT:    v_readfirstlane_b32 s15, v3
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readfirstlane_b32 s12, v9
; GFX9-NEXT:    v_readfirstlane_b32 s14, v8
; GFX9-NEXT:    v_readfirstlane_b32 s47, v7
; GFX9-NEXT:    v_readfirstlane_b32 s59, v6
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; GFX9-NEXT:    v_readfirstlane_b32 s42, v13
; GFX9-NEXT:    v_readfirstlane_b32 s74, v30
; GFX9-NEXT:    v_readfirstlane_b32 s44, v29
; GFX9-NEXT:    v_readfirstlane_b32 s56, v28
; GFX9-NEXT:    v_readfirstlane_b32 s62, v27
; GFX9-NEXT:    v_readfirstlane_b32 s72, v26
; GFX9-NEXT:    v_readfirstlane_b32 s13, v25
; GFX9-NEXT:    v_readfirstlane_b32 s40, v24
; GFX9-NEXT:    v_readfirstlane_b32 s57, v23
; GFX9-NEXT:    v_readfirstlane_b32 s60, v22
; GFX9-NEXT:    v_readfirstlane_b32 s9, v21
; GFX9-NEXT:    v_readfirstlane_b32 s11, v20
; GFX9-NEXT:    v_readfirstlane_b32 s41, v19
; GFX9-NEXT:    v_readfirstlane_b32 s76, v18
; GFX9-NEXT:    v_readfirstlane_b32 s45, v17
; GFX9-NEXT:    v_readfirstlane_b32 s58, v16
; GFX9-NEXT:    v_readfirstlane_b32 s73, v15
; GFX9-NEXT:    v_readfirstlane_b32 s75, v14
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:312
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:296
; GFX9-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:292
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:280
; GFX9-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:276
; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:272
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:268
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:264
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:260
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:256
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:252
; GFX9-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:248
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:244
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:184
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:180
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:176
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:172
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:168
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:164
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:12
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    s_cbranch_scc0 .LBB89_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_mov_b32_e32 v8, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v8
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v8
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v8
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v8
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v8
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v8
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v8
; GFX9-NEXT:    v_mov_b32_e32 v4, s6
; GFX9-NEXT:    v_perm_b32 v4, s7, v4, v8
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX9-NEXT:    v_mov_b32_e32 v4, s15
; GFX9-NEXT:    v_perm_b32 v4, s43, v4, v8
; GFX9-NEXT:    v_mov_b32_e32 v5, s8
; GFX9-NEXT:    v_perm_b32 v5, s10, v5, v8
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX9-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX9-NEXT:    v_mov_b32_e32 v5, s47
; GFX9-NEXT:    v_perm_b32 v5, s59, v5, v8
; GFX9-NEXT:    v_mov_b32_e32 v6, s12
; GFX9-NEXT:    v_perm_b32 v6, s14, v6, v8
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX9-NEXT:    v_mov_b32_e32 v6, s61
; GFX9-NEXT:    v_perm_b32 v6, s63, v6, v8
; GFX9-NEXT:    v_mov_b32_e32 v7, s42
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v8
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v7, s73
; GFX9-NEXT:    v_perm_b32 v7, s75, v7, v8
; GFX9-NEXT:    v_mov_b32_e32 v63, v33
; GFX9-NEXT:    v_mov_b32_e32 v33, v9
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_perm_b32 v9, s58, v9, v8
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX9-NEXT:    v_lshl_or_b32 v7, v9, 16, v7
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v10
; GFX9-NEXT:    v_perm_b32 v10, s74, v12, v8
; GFX9-NEXT:    v_or_b32_sdwa v9, v11, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v11, v9, 16, v10
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v18
; GFX9-NEXT:    v_or_b32_sdwa v9, v22, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v13
; GFX9-NEXT:    v_or_b32_sdwa v10, v15, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_lshl_or_b32 v12, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v14
; GFX9-NEXT:    v_perm_b32 v10, v26, v21, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v17, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v13, v9, 16, v10
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
; GFX9-NEXT:    v_or_b32_sdwa v9, v30, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v16
; GFX9-NEXT:    v_or_b32_sdwa v10, v20, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v14, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v19
; GFX9-NEXT:    v_perm_b32 v10, v35, v29, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v24, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v15, v9, 16, v10
; GFX9-NEXT:    v_perm_b32 v9, v39, v34, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v51, v36, s4
; GFX9-NEXT:    v_perm_b32 v16, v28, v23, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_perm_b32 v17, v37, v27, s4
; GFX9-NEXT:    v_lshl_or_b32 v16, v16, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v41, v50, s4
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v10
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v52, v31, s4
; GFX9-NEXT:    v_lshl_or_b32 v18, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v40, v46, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v32, v38, s4
; GFX9-NEXT:    v_lshl_or_b32 v19, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v55, v57, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v45, v49, s4
; GFX9-NEXT:    v_lshl_or_b32 v20, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v53, v60, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v56, v44, s4
; GFX9-NEXT:    v_lshl_or_b32 v21, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v58, v42, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v59, v47, s4
; GFX9-NEXT:    v_lshl_or_b32 v22, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v50, v32
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v51, v49
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v41, v46
; GFX9-NEXT:    v_mov_b32_e32 v39, v52
; GFX9-NEXT:    v_mov_b32_e32 v52, v38
; GFX9-NEXT:    v_mov_b32_e32 v34, v37
; GFX9-NEXT:    v_mov_b32_e32 v37, v31
; GFX9-NEXT:    s_lshl_b32 s5, s13, 8
; GFX9-NEXT:    s_lshl_b32 s77, s9, 8
; GFX9-NEXT:    s_lshl_b32 s78, s44, 8
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_perm_b32 v9, v9, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_perm_b32 v10, v32, v58, s4
; GFX9-NEXT:    v_lshl_or_b32 v23, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v49
; GFX9-NEXT:    v_or_b32_sdwa v10, v24, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v24, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v35
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v25, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v25, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v33
; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v33, v46
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    v_or_b32_sdwa v9, v43, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v26, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v48
; GFX9-NEXT:    v_perm_b32 v10, v63, v61, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v27, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v54, v61
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_mov_b32_e32 v48, v63
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
; GFX9-NEXT:    v_or_b32_sdwa v10, v28, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshl_or_b32 v28, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v36, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v29, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v43
; GFX9-NEXT:    v_or_b32_sdwa v9, v62, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
; GFX9-NEXT:    v_or_b32_sdwa v10, v30, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshl_or_b32 v30, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
; GFX9-NEXT:    v_perm_b32 v10, v63, v61, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v46, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v31, v9, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v9, s57
; GFX9-NEXT:    v_perm_b32 v8, s60, v9, v8
; GFX9-NEXT:    s_and_b32 s4, s40, 0xff
; GFX9-NEXT:    s_or_b32 s4, s4, s5
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX9-NEXT:    v_lshl_or_b32 v9, s4, 16, v8
; GFX9-NEXT:    s_and_b32 s4, s76, 0xff
; GFX9-NEXT:    s_lshl_b32 s5, s41, 8
; GFX9-NEXT:    s_or_b32 s4, s4, s5
; GFX9-NEXT:    s_and_b32 s5, s11, 0xff
; GFX9-NEXT:    s_or_b32 s5, s5, s77
; GFX9-NEXT:    s_pack_ll_b32_b16 s4, s4, s5
; GFX9-NEXT:    s_and_b32 s5, s72, 0xff
; GFX9-NEXT:    s_lshl_b32 s77, s62, 8
; GFX9-NEXT:    s_or_b32 s5, s5, s77
; GFX9-NEXT:    s_and_b32 s77, s56, 0xff
; GFX9-NEXT:    s_or_b32 s77, s77, s78
; GFX9-NEXT:    s_pack_ll_b32_b16 s5, s5, s77
; GFX9-NEXT:    v_mov_b32_e32 v8, s4
; GFX9-NEXT:    v_mov_b32_e32 v10, s5
; GFX9-NEXT:    s_mov_b64 s[4:5], 0
; GFX9-NEXT:    v_mov_b32_e32 v63, v36
; GFX9-NEXT:    s_branch .LBB89_3
; GFX9-NEXT:  .LBB89_2:
; GFX9-NEXT:    v_mov_b32_e32 v54, v61
; GFX9-NEXT:    v_mov_b32_e32 v50, v32
; GFX9-NEXT:    v_mov_b32_e32 v51, v49
; GFX9-NEXT:    v_mov_b32_e32 v39, v52
; GFX9-NEXT:    v_mov_b32_e32 v52, v38
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v41, v46
; GFX9-NEXT:    v_mov_b32_e32 v34, v37
; GFX9-NEXT:    v_mov_b32_e32 v37, v31
; GFX9-NEXT:    s_mov_b64 s[4:5], -1
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:  .LBB89_3: ; %Flow
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_vccnz .LBB89_5
; GFX9-NEXT:  ; %bb.4: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v48
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v0, v0, v61, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v33
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v0, v0, v38, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v62
; GFX9-NEXT:    v_perm_b32 v0, v0, v43, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    s_add_i32 s74, s74, 3
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    s_add_i32 s72, s72, 3
; GFX9-NEXT:    s_add_i32 s56, s56, 3
; GFX9-NEXT:    s_add_i32 s60, s60, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s76, s76, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s75, s75, 3
; GFX9-NEXT:    s_add_i32 s58, s58, 3
; GFX9-NEXT:    s_add_i32 s63, s63, 3
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s59, s59, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    v_mov_b32_e32 v33, s17
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v63
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v38, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v62, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v54, s4
; GFX9-NEXT:    v_add_u32_e32 v63, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v30, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v35, s4
; GFX9-NEXT:    v_add_u32_e32 v29, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v36
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v28, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v49, s4
; GFX9-NEXT:    v_add_u32_e32 v27, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v26, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v32
; GFX9-NEXT:    v_perm_b32 v0, v0, v58, s4
; GFX9-NEXT:    v_add_u32_e32 v25, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v32, s23
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v42, s4
; GFX9-NEXT:    v_add_u32_e32 v24, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v59
; GFX9-NEXT:    v_perm_b32 v0, v0, v47, s4
; GFX9-NEXT:    v_add_u32_e32 v23, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v53
; GFX9-NEXT:    v_perm_b32 v0, v0, v60, s4
; GFX9-NEXT:    v_add_u32_e32 v22, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v56
; GFX9-NEXT:    v_perm_b32 v0, v0, v44, s4
; GFX9-NEXT:    v_add_u32_e32 v21, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v55
; GFX9-NEXT:    v_perm_b32 v0, v0, v57, s4
; GFX9-NEXT:    v_add_u32_e32 v20, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v45
; GFX9-NEXT:    v_perm_b32 v0, v0, v51, s4
; GFX9-NEXT:    v_add_u32_e32 v49, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v40
; GFX9-NEXT:    v_perm_b32 v0, v0, v41, s4
; GFX9-NEXT:    v_add_u32_e32 v19, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v50
; GFX9-NEXT:    v_perm_b32 v0, v0, v52, s4
; GFX9-NEXT:    v_add_u32_e32 v52, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX9-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff, v24
; GFX9-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff, v26
; GFX9-NEXT:    v_lshl_or_b32 v23, v25, 16, v23
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff, v30
; GFX9-NEXT:    v_lshl_or_b32 v25, v29, 16, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff, v28
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v24, v27, 16, v24
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff, v63
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX9-NEXT:    v_lshl_or_b32 v19, v52, 16, v19
; GFX9-NEXT:    v_lshl_or_b32 v20, v49, 16, v20
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v18, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v39
; GFX9-NEXT:    v_perm_b32 v0, v0, v37, s4
; GFX9-NEXT:    v_add_u32_e32 v37, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX9-NEXT:    v_lshl_or_b32 v18, v37, 16, v18
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v54, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v34
; GFX9-NEXT:    v_mov_b32_e32 v34, s19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v17, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v16, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v55, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v16, v55, 16, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v15, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v40, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v15, v40, 16, v15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v14, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v41, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v14, v41, 16, v14
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v13, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v42, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v13, v42, 16, v13
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v12, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v43, 0x300, v0
; GFX9-NEXT:    v_mov_b32_e32 v0, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v32, s22, v32, v0
; GFX9-NEXT:    v_add_u32_e32 v32, 0x300, v32
; GFX9-NEXT:    v_perm_b32 v33, s16, v33, v0
; GFX9-NEXT:    v_add_u32_e32 v33, 0x300, v33
; GFX9-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX9-NEXT:    v_lshl_or_b32 v12, v43, 16, v12
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, s74, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v11, 0x300, v1
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_add_u32_e32 v44, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s62
; GFX9-NEXT:    v_perm_b32 v1, s72, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v10, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s44
; GFX9-NEXT:    v_perm_b32 v1, s56, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v45, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s57
; GFX9-NEXT:    v_perm_b32 v1, s60, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v9, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s13
; GFX9-NEXT:    v_perm_b32 v1, s40, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v46, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s41
; GFX9-NEXT:    v_perm_b32 v1, s76, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v8, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s9
; GFX9-NEXT:    v_perm_b32 v1, s11, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v47, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s73
; GFX9-NEXT:    v_perm_b32 v1, s75, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v7, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s45
; GFX9-NEXT:    v_perm_b32 v1, s58, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v56, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s61
; GFX9-NEXT:    v_perm_b32 v1, s63, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v6, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s42
; GFX9-NEXT:    v_perm_b32 v1, s46, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v57, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s47
; GFX9-NEXT:    v_perm_b32 v1, s59, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v5, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s12
; GFX9-NEXT:    v_perm_b32 v1, s14, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v58, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s15
; GFX9-NEXT:    v_perm_b32 v1, s43, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v4, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    v_perm_b32 v1, s10, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v59, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s29
; GFX9-NEXT:    v_perm_b32 v1, s28, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v3, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    v_perm_b32 v1, s7, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v60, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s25
; GFX9-NEXT:    v_perm_b32 v1, s24, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s27
; GFX9-NEXT:    v_perm_b32 v1, s26, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v61, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s21
; GFX9-NEXT:    v_perm_b32 v1, s20, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX9-NEXT:    v_lshl_or_b32 v1, v32, 16, v1
; GFX9-NEXT:    v_and_b32_e32 v32, 0xffff, v54
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v32
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX9-NEXT:    v_lshl_or_b32 v29, v30, 16, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v26, v31, 16, v26
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, s18, v34, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff, v38
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v2, v61, 16, v2
; GFX9-NEXT:    v_lshl_or_b32 v3, v60, 16, v3
; GFX9-NEXT:    v_lshl_or_b32 v4, v59, 16, v4
; GFX9-NEXT:    v_lshl_or_b32 v5, v58, 16, v5
; GFX9-NEXT:    v_lshl_or_b32 v6, v57, 16, v6
; GFX9-NEXT:    v_lshl_or_b32 v7, v56, 16, v7
; GFX9-NEXT:    v_lshl_or_b32 v8, v47, 16, v8
; GFX9-NEXT:    v_lshl_or_b32 v9, v46, 16, v9
; GFX9-NEXT:    v_lshl_or_b32 v10, v45, 16, v10
; GFX9-NEXT:    v_lshl_or_b32 v11, v44, 16, v11
; GFX9-NEXT:    v_lshl_or_b32 v28, v62, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshl_or_b32 v30, v31, 16, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX9-NEXT:    v_lshl_or_b32 v31, v32, 16, v31
; GFX9-NEXT:  .LBB89_5: ; %end
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64bf16_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:260
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:188
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:60
; GFX11-TRUE16-NEXT:    s_clause 0xf
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s74, v30
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v29
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v28
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v27
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v26
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v25
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v24
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v23
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v22
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v21
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v20
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v19
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v18
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v17
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v16
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s75, 0
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB89_4
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, 0xc0c0004 :: v_dual_lshlrev_b32 v13, 8, v86
; GFX11-TRUE16-NEXT:    s_and_b32 s76, s43, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s77, s11, 8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s2, s3, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s22, s23, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s26, s27, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s0, s1, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s18, s19, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s9, s4, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s24, s25, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s40, s8, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s14, s7, v8
; GFX11-TRUE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s16, s17, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s12, s6, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v4.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s28, s29, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s76 :: v_dual_lshlrev_b32 v15, 8, v70
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s61, s45, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s60, s15, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v7.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s72, s58, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v9.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s59, s44, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v68
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s41, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s89, s10, 8
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s73, s42, v8
; GFX11-TRUE16-NEXT:    s_or_b32 s88, s88, s89
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v12
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s88
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s74, 0xff
; GFX11-TRUE16-NEXT:    s_and_b32 s77, s63, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s76, s57, 8
; GFX11-TRUE16-NEXT:    s_and_b32 s78, s56, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s79, s5, 8
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, s88, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v38
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v32
; GFX11-TRUE16-NEXT:    s_or_b32 s76, s77, s76
; GFX11-TRUE16-NEXT:    s_or_b32 s77, s78, s79
; GFX11-TRUE16-NEXT:    s_and_b32 s78, s62, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s79, s47, 8
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s46, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s89, s13, 8
; GFX11-TRUE16-NEXT:    s_or_b32 s78, s78, s79
; GFX11-TRUE16-NEXT:    s_or_b32 s79, s88, s89
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v8.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v14, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v83
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v33
; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s77
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v50
; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s78, s79
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v39
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v134, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v8.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v15, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v53, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v99, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v66, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v97, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v71, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v103, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v69, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v116, v67, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v84, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v130, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v98, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v145, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v113, v55, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v135, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v147, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v118, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v117
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v64
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v151, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v164
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v119
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v131
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v82
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v149
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v8, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v176, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v26, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v144
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v87
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v128
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v167
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v101
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v160
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v102
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v8, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v28, v29
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v162
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v132
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v177, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v179
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v150
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v133
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v40, 0xff, v182
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v8, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v166
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v165
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v129
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v41, v29, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v178
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v181, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v183, v8, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s76
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v40, v31
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v180, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s77
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v30.h, v183.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v41.l
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB89_3
; GFX11-TRUE16-NEXT:  .LBB89_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v180
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v166
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v182
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v165
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v181
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s59, s59, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v160
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v179
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v162
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v1, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v177
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v102, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v149
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v132, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v167
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v144
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v176
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v101, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v131
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v164
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v87, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v117
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v151
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v118
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v82, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v147
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v113
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v85, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v135
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v55, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v98
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v145
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v55, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v130
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v52, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v69
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v116
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v71
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v67, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v103
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v66
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v97
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v53
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v99
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v50
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v38
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v134
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v114
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v68
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s74, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s73, s73, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s73, s42, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s62, s47, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s59, s44, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s43, s43, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s72, s58, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s43, s11, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s56, s5, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s12, s6, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s61, s61, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s41, s10, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s46, s13, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s60, s15, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s61, s45, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v65, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s28, s29, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s63, s63, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s63, s57, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v48, 0x300, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s9, s4, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s20, s21, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s14, s7, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v54, s40, s8, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v67, s26, s27, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v68, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s22, s23, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v69, s2, s3, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v70, s18, s19, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v66, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s24, s25, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v71, s16, s17, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v80, s0, s1, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v54, 0x300, v54
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v67, 0x300, v67
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v69
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v70
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v81, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v71
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v80
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v69.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v70.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v81.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v67.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v68.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v54.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v66.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v65.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v53.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v50.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v48.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v39.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v37.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v38.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v32.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v33.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v34.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v35.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v51.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v36.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v49.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v52.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v55.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v85.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v64.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v82.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v87.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v101.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v28.h, v132.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v102.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v30.h, v129.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v31.h, v133.l
; GFX11-TRUE16-NEXT:  .LBB89_3: ; %end
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT:  .LBB89_4:
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_branch .LBB89_2
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64bf16_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:260
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:188
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:132
; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v54, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v52, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v55, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:60
; GFX11-FAKE16-NEXT:    s_clause 0xf
; GFX11-FAKE16-NEXT:    scratch_load_u16 v34, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v50, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v39, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v36, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v49, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v37, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v53, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v32, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v48, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v51, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v38, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v35, off, s32
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s74, v30
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v29
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v28
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v27
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v26
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v25
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v24
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v23
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v22
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v21
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v20
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v19
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v18
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v17
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v16
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s75, 0
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:320 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-FAKE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB89_4
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s74, 0xff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v48
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v32
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s0, s1, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s16, s17, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s20, s21, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s2, s3, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s18, s19, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s22, s23, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s24, s25, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s28, s29, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s11, s8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s6, s4, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s7, s5, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v67, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v80, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v68, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v10, 16, v11
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s26, s27, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s42, s15, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s44, s41, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v69, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v7, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s12, s9, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v116, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_lshl_b32 s77, s43, 8
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v118, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s14, s10, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, s76, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v38
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s46, 0xff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v53
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v9, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v49
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v36
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v11, 16, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v14, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v50
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v65, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v14, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v82, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v115, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v130
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v9, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v96, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v71, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v98, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v86, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v117
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v144
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v102, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s60, s57, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v87, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v135
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v150
; GFX11-FAKE16-NEXT:    s_lshl_b32 s77, s47, 8
; GFX11-FAKE16-NEXT:    s_and_b32 s78, s40, 0xff
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v11, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v112, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, s76, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v132, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s58, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s79, s13, 8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v11, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v113, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-FAKE16-NEXT:    s_or_b32 s77, s78, s79
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s73, s72, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v11, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v134
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v119
; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s77
; GFX11-FAKE16-NEXT:    s_and_b32 s77, s62, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s78, s59, 8
; GFX11-FAKE16-NEXT:    s_lshl_b32 s79, s45, 8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v24, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v131
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v114
; GFX11-FAKE16-NEXT:    s_or_b32 s77, s77, s78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_and_b32 s78, s56, 0xff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v24, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v146, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_or_b32 s78, s78, s79
; GFX11-FAKE16-NEXT:    s_and_b32 s79, s63, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s88, s61, 8
; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v160
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v147
; GFX11-FAKE16-NEXT:    s_or_b32 s78, s79, s88
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v148
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v26, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v149
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v133
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v26, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v162, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v177
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v164
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v151
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v28, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v176
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v161
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v28, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v179, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v181
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v165
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v178
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v180
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v166
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v183, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, s77
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v11, v31
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v182, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v183
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v40, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, s78, 16, v8
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, s76
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v40
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB89_3
; GFX11-FAKE16-NEXT:  .LBB89_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v182
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v180
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v178
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v179
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v176
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v177
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v3, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v4, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v5, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v151
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v150, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v151, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v162
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v149
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v148
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v132
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v135, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v160
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v131
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v144, 0x300, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v134
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v115
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v115, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v118
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v130
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v113
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v112
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v114, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v116
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v87
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v96
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v100, 0x300, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v98
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v86
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v71
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v71, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v82
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v68
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v69
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v80
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v54, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v65
; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v52, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v55, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v53
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v49
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v38
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v2
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v48
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s74, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s73, s73, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s62, s59, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s63, s63, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s73, s72, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s40, s13, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s63, s61, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s12, s9, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s56, s45, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s60, s57, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s46, s43, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s58, s47, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s11, s8, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v50, s24, s25, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, s0, s1, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s44, s41, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s14, s10, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s42, s15, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s26, s27, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v67, s2, s3, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v65, 0x300, v65
; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s7, s5, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s28, s29, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v64, s16, s17, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v67, 0x300, v67
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff, v65
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s6, s4, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s20, s21, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v66, s18, s19, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v64
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v67, 16, v65
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v69, 16, v50
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v38
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v32, 16, v36
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v34, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v54
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v17
; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s22, s23, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v66, 0x300, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v64
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v68, 0x300, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v51
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v55, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v52, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v99
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v115
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v66, 16, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v68
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v53, 16, v51
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v100, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v97, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v135
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v151
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v64
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v65
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v49, 16, v48
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v50
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v51
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v38
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v70, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v114, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v144, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v133, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v150, 16, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v35
; GFX11-FAKE16-NEXT:  .LBB89_3: ; %end
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:320 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-FAKE16-NEXT:  .LBB89_4:
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_branch .LBB89_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) {
; SI-LABEL: bitcast_v64bf16_to_v128i8:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:8
; SI-NEXT:    v_mov_b32_e32 v31, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v30
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v62
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v3
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v4
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v56
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_and_b32_e32 v46, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v15
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v33
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v36
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v39
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v39
; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v36
; SI-NEXT:    v_and_b32_e32 v36, 0xffff0000, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v36
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v33
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v7
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v33
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v53
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v19
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v51
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    v_mul_f32_e32 v51, 1.0, v63
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v58
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; kill: killed $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v15
; SI-NEXT:    v_mul_f32_e32 v49, 1.0, v49
; SI-NEXT:    v_mul_f32_e32 v53, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v43
; SI-NEXT:    v_mul_f32_e32 v55, 1.0, v55
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v42
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v41
; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v40
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v41, 1.0, v23
; SI-NEXT:    v_mul_f32_e32 v33, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v25
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v38
; SI-NEXT:    v_mul_f32_e32 v45, 1.0, v37
; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v35
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v34
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v29
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v60
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v59
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; kill: killed $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; kill: killed $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB90_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_alignbit_b32 v38, v0, v1, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v35, v0, v1, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v32, v0, v1, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v29, v0, v1, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v26, v0, v1, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_alignbit_b32 v40, v0, v1, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v23, v0, v1, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_alignbit_b32 v54, v0, v1, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v18, v0, v1, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_alignbit_b32 v52, v0, v1, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_alignbit_b32 v15, v0, v27, 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v50, v0, v1, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v49
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v13, v0, v53, 16
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v49
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_alignbit_b32 v48, v0, v1, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v55
; SI-NEXT:    v_alignbit_b32 v11, v0, v19, 16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v53
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 24, v6
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 24, v51
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v40, v38, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v40, v38, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v40, v38, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v54, v35, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v54, v35, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v54, v35, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v52, v32, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v52, v32, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v52, v32, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v50, v29, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v50, v29, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v50, v29, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v48, v26, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v48, v26, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v48, v26, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_alignbit_b32 v37, v0, v1, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v62
; SI-NEXT:    v_alignbit_b32 v8, v0, v21, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v37, v23, 24
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v55
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v37, v23, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v37, v23, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v21, 24, v59
; SI-NEXT:    v_mov_b32_e32 v59, v19
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_alignbit_b32 v34, v0, v1, 16
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v42
; SI-NEXT:    v_alignbit_b32 v5, v0, v41, 16
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v6, v34, v18, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v34, v18, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v34, v18, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4) expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v57
; SI-NEXT:    v_alignbit_b32 v4, v1, v36, 16
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v43
; SI-NEXT:    v_alignbit_b32 v28, v1, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v45
; SI-NEXT:    v_alignbit_b32 v3, v1, v61, 16
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v6, v0, v15, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v0, v15, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v0, v15, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v28, v13, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v28, v13, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v28, v13, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v9
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v56
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v33
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v12
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_alignbit_b32 v25, v1, v2, 16
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v44
; SI-NEXT:    v_alignbit_b32 v2, v1, v47, 16
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v56
; SI-NEXT:    v_alignbit_b32 v22, v1, v7, 16
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v33
; SI-NEXT:    v_alignbit_b32 v20, v7, v10, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v6, v25, v11, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v25, v11, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v25, v11, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v22, v8, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v22, v8, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v22, v8, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v20, v5, 24
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v20, v5, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v20, v5, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v46
; SI-NEXT:    v_alignbit_b32 v1, v1, v60, 16
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_alignbit_b32 v17, v7, v10, 16
; SI-NEXT:    v_alignbit_b32 v6, v17, v4, 24
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v12
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v17, v4, 16
; SI-NEXT:    v_alignbit_b32 v14, v7, v24, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v17, v4, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v14, v3, 16
; SI-NEXT:    v_alignbit_b32 v10, v19, v30, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v14, v3, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v10, v2, 24
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v51
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v10, v2, 16
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v7, v7, v63, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v10, v2, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v7, v1, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v7, v1, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v40
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v54
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v52
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v50
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v48
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v37
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v34
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v0
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v28
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v25
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v22
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v20
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v17
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v14
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v10
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v7
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v58
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v39
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v27
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v49
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v53
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v55
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v41
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v43
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v39, v14, v3, 24
; SI-NEXT:    v_alignbit_b32 v58, v7, v1, 8
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:  .LBB90_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB90_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v46
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v60
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v44
; SI-NEXT:    v_alignbit_b32 v1, v1, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v45
; SI-NEXT:    v_alignbit_b32 v2, v2, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v61
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v57
; SI-NEXT:    v_alignbit_b32 v3, v3, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v51
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v42
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_add_f32_e32 v36, 0x40c00000, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_alignbit_b32 v4, v4, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v63
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_alignbit_b32 v5, v7, v5, 16
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v36
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v62
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v7, v7, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v12
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v55
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v6
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v19
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v8, v10, v8, 16
; SI-NEXT:    v_alignbit_b32 v10, v59, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v24
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_alignbit_b32 v11, v13, v11, 16
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; SI-NEXT:    v_alignbit_b32 v14, v13, v0, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_alignbit_b32 v13, v15, v13, 16
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v16
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_lshrrev_b32_e32 v6, 24, v6
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 24, v36
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v58, v7, v1, 8
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_alignbit_b32 v17, v15, v0, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v33
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v27
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_alignbit_b32 v15, v18, v15, 16
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v21
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; SI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_alignbit_b32 v29, v30, v29, 16
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v34
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; SI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_alignbit_b32 v32, v34, v32, 16
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v38
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v35
; SI-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; SI-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_alignbit_b32 v35, v38, v35, 16
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v49
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v50
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v52
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v54
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; SI-NEXT:    v_add_f32_e32 v51, 0x40c00000, v50
; SI-NEXT:    v_add_f32_e32 v53, 0x40c00000, v52
; SI-NEXT:    v_add_f32_e32 v55, 0x40c00000, v54
; SI-NEXT:    v_add_f32_e32 v41, 0x40c00000, v40
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v51
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v41
; SI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_alignbit_b32 v20, v18, v0, 16
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v56
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_alignbit_b32 v18, v19, v18, 16
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v27
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v22, v19, v0, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_alignbit_b32 v23, v23, v19, 16
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v9
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v25, v19, v0, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v9
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v27
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v21
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v9, 24, v12
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_alignbit_b32 v26, v26, v19, 16
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v24
; SI-NEXT:    v_alignbit_b32 v28, v19, v0, 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v19
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v0, v30, v0, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v19
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v24
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v34, v34, v33, 16
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v30, 24, v30
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; SI-NEXT:    v_add_f32_e32 v37, 0x40c00000, v33
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v33
; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v33
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v37, v38, v37, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v33
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v48, 0x40c00000, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v39, 0x40c00000, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v38
; SI-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; SI-NEXT:    v_alignbit_b32 v38, v49, v38, 16
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v39
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v48, v49, v48, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v39, 24, v39
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v39, v14, v3, 24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; SI-NEXT:    v_alignbit_b32 v50, v50, v49, 16
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; SI-NEXT:    v_alignbit_b32 v52, v52, v49, 16
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; SI-NEXT:    v_alignbit_b32 v54, v54, v49, 16
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v49
; SI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; SI-NEXT:    v_alignbit_b32 v40, v40, v49, 16
; SI-NEXT:    v_alignbit_b32 v6, v40, v38, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v40, v38, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v40, v38, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v54, v35, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v54, v35, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v54, v35, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v52, v32, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v52, v32, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v52, v32, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v50, v29, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v50, v29, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v50, v29, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v48, v26, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v48, v26, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v48, v26, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v37, v23, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v37, v23, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v37, v23, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v34, v18, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v34, v18, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v34, v18, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v0, v15, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v0, v15, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v0, v15, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v28, v13, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v28, v13, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v28, v13, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v25, v11, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v25, v11, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v25, v11, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v22, v8, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v22, v8, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v22, v8, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v20, v5, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v20, v5, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v20, v5, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v17, v4, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v17, v4, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v17, v4, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v14, v3, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v14, v3, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v10, v2, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v10, v2, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v10, v2, 8
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v7, v1, 24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v7, v1, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v40
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v54
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v52
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v50
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v48
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v37
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v34
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v0
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v28
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v25
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v22
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v20
; SI-NEXT:    v_lshrrev_b32_e32 v49, 24, v41
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v17
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 24, v55
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v14
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 24, v53
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v10
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 24, v51
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 8, v7
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:  .LBB90_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v38
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_store_dword v6, v31, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v40
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 4, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v35
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 8, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v54
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 12, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v32
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 16, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v52
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 20, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v29
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 24, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v50
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 28, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v26
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 32, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v48
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 36, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v23
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 40, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v37
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 44, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v18
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 48, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v34
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 52, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v15
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v12, 24, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v12, v9
; SI-NEXT:    v_or_b32_e32 v6, v6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 56, v31
; SI-NEXT:    buffer_store_dword v6, v9, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v9, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 60, v31
; SI-NEXT:    buffer_store_dword v0, v6, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v13
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v9, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 64, v31
; SI-NEXT:    buffer_store_dword v0, v6, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v28
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v9, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x44, v31
; SI-NEXT:    buffer_store_dword v0, v6, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v11
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v9, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x48, v31
; SI-NEXT:    buffer_store_dword v0, v6, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v25
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v9, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x4c, v31
; SI-NEXT:    buffer_store_dword v0, v6, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v8
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v8, 24, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x50, v31
; SI-NEXT:    buffer_store_dword v0, v6, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v22
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v8, 24, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v8, v6
; SI-NEXT:    v_or_b32_e32 v0, v0, v6
; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x54, v31
; SI-NEXT:    buffer_store_dword v0, v6, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, 0x58, v31
; SI-NEXT:    buffer_store_dword v0, v5, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v20
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_or_b32_e32 v0, v0, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, 0x5c, v31
; SI-NEXT:    buffer_store_dword v0, v5, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x60, v31
; SI-NEXT:    buffer_store_dword v0, v4, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v17
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_or_b32_e32 v0, v0, v4
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x64, v31
; SI-NEXT:    buffer_store_dword v0, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v39
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x68, v31
; SI-NEXT:    buffer_store_dword v0, v3, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v14
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v4
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_or_b32_e32 v0, v0, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x6c, v31
; SI-NEXT:    buffer_store_dword v0, v3, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v31
; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v10
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v59
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v31
; SI-NEXT:    buffer_store_dword v0, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v58
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x78, v31
; SI-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v7
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v2
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x7c, v31
; SI-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v128i8:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; kill: killed $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB90_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[7:8]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    v_lshrrev_b32_e32 v40, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, v53
; VI-NEXT:    v_mov_b32_e32 v53, v40
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[17:18]
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v45, 8, v29
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v51, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v56, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v54, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v48, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v55, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v43, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v61, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v36, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v49, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v30
; VI-NEXT:    v_lshrrev_b32_e32 v42, 16, v29
; VI-NEXT:    v_lshrrev_b32_e32 v57, 24, v28
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v28
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v27
; VI-NEXT:    v_lshrrev_b32_e32 v50, 8, v27
; VI-NEXT:    v_lshrrev_b32_e32 v41, 16, v17
; VI-NEXT:    v_lshrrev_b32_e32 v60, 8, v17
; VI-NEXT:    v_mov_b32_e32 v38, v45
; VI-NEXT:    v_lshrrev_b32_e32 v40, 8, v1
; VI-NEXT:    v_mov_b32_e32 v39, v46
; VI-NEXT:  .LBB90_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB90_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v18
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, 0x7fff, v34
; VI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v18, 16, 1
; VI-NEXT:    s_movk_i32 s6, 0x7fff
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v18
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_alignbit_b32 v18, v18, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v17
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    v_cndmask_b32_e32 v17, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_alignbit_b32 v17, v17, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v20
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v20, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v20
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; VI-NEXT:    v_cndmask_b32_e32 v20, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_alignbit_b32 v20, v20, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v19
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v19
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_cndmask_b32_e32 v19, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_alignbit_b32 v19, v19, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v22
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v22, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v22
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v22
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; VI-NEXT:    v_cndmask_b32_e32 v22, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_alignbit_b32 v22, v22, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v21
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v21
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_cndmask_b32_e32 v21, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_alignbit_b32 v21, v21, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v24
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v24, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v24
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v24
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; VI-NEXT:    v_cndmask_b32_e32 v24, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_alignbit_b32 v24, v24, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v23
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v23
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_cndmask_b32_e32 v23, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_alignbit_b32 v23, v23, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v26
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v26, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v26
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v26
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; VI-NEXT:    v_cndmask_b32_e32 v26, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_alignbit_b32 v26, v26, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v25
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v25
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_cndmask_b32_e32 v25, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_alignbit_b32 v25, v25, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v28
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v28, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v28
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v28
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; VI-NEXT:    v_cndmask_b32_e32 v28, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_alignbit_b32 v28, v28, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v27
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v27
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_cndmask_b32_e32 v27, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_alignbit_b32 v27, v27, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v30
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v30, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v30
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v30
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; VI-NEXT:    v_cndmask_b32_e32 v30, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_alignbit_b32 v30, v30, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v29
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v29, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v29
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_cndmask_b32_e32 v29, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_alignbit_b32 v29, v29, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v32
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v32
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v32
; VI-NEXT:    v_alignbit_b32 v32, v32, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v31
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v31
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_alignbit_b32 v31, v31, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v2
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v2
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_alignbit_b32 v2, v2, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v1
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v1
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_alignbit_b32 v1, v1, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v4
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v4, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v4
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v4
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; VI-NEXT:    v_cndmask_b32_e32 v4, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_alignbit_b32 v4, v4, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v3
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v3
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_alignbit_b32 v3, v3, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v6
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v6, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v6
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v6
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; VI-NEXT:    v_cndmask_b32_e32 v6, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_alignbit_b32 v6, v6, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v5
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v5
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_alignbit_b32 v5, v5, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v8
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v8, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v8
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v8
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; VI-NEXT:    v_cndmask_b32_e32 v8, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_alignbit_b32 v8, v8, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v7
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v7
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_alignbit_b32 v7, v7, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v10
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v10, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v10
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v10
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; VI-NEXT:    v_cndmask_b32_e32 v10, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_alignbit_b32 v10, v10, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v9
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v9
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_alignbit_b32 v9, v9, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v12
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v12, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v12
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v12
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; VI-NEXT:    v_cndmask_b32_e32 v12, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_alignbit_b32 v12, v12, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v11
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v11
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_alignbit_b32 v11, v11, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v14
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v14, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v14
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v14
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; VI-NEXT:    v_cndmask_b32_e32 v14, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_alignbit_b32 v14, v14, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v13
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v13
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_alignbit_b32 v13, v13, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v16
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v16
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_alignbit_b32 v16, v16, v33, 16
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v15
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v15
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v34, v35, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_alignbit_b32 v15, v15, v33, 16
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[7:8]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[1:2]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v32
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v31
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v30
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v26
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v25
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v23
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v22
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v21
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v20
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v19
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v18
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v18
; VI-NEXT:    v_lshrrev_b32_e32 v63, 8, v9
; VI-NEXT:    v_lshrrev_b32_e32 v47, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v51, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v56, 8, v8
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v7
; VI-NEXT:    v_lshrrev_b32_e32 v54, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v48, 16, v5
; VI-NEXT:    v_lshrrev_b32_e32 v55, 8, v5
; VI-NEXT:    v_lshrrev_b32_e32 v37, 24, v4
; VI-NEXT:    v_lshrrev_b32_e32 v43, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v61, 8, v4
; VI-NEXT:    v_lshrrev_b32_e32 v36, 16, v3
; VI-NEXT:    v_lshrrev_b32_e32 v62, 8, v3
; VI-NEXT:    v_lshrrev_b32_e32 v49, 24, v2
; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v39, 16, v1
; VI-NEXT:    v_lshrrev_b32_e32 v40, 8, v1
; VI-NEXT:    v_lshrrev_b32_e32 v52, 8, v30
; VI-NEXT:    v_lshrrev_b32_e32 v42, 16, v29
; VI-NEXT:    v_lshrrev_b32_e32 v38, 8, v29
; VI-NEXT:    v_lshrrev_b32_e32 v57, 24, v28
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v28
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v28
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v27
; VI-NEXT:    v_lshrrev_b32_e32 v50, 8, v27
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v18
; VI-NEXT:    v_lshrrev_b32_e32 v41, 16, v17
; VI-NEXT:    v_lshrrev_b32_e32 v60, 8, v17
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:  .LBB90_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v1, v1, v40, s4
; VI-NEXT:    v_perm_b32 v2, v2, v35, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v45, v39, v45, s4
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v45
; VI-NEXT:    v_or_b32_e32 v1, v1, v33
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v1, v53, v49, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v62, s4
; VI-NEXT:    v_perm_b32 v1, v36, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v43, v37, s4
; VI-NEXT:    v_perm_b32 v1, v4, v61, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v5, v55, s4
; VI-NEXT:    v_perm_b32 v1, v48, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v6, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v7, v54, s4
; VI-NEXT:    v_perm_b32 v1, v44, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v51, v47, s4
; VI-NEXT:    v_perm_b32 v1, v8, v56, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    v_perm_b32 v2, v9, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v10, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v11, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v12, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v13, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v14, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v15, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v16, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v17, v60, s4
; VI-NEXT:    v_perm_b32 v1, v41, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v18, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v19, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v20, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v21, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v22, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v23, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v24, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v25, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v26, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v27, v50, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v59, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v34, v57, s4
; VI-NEXT:    v_perm_b32 v1, v28, v58, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v29, v38, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v42, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v30, v52, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v31, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v32, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v128i8:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32
; GFX9-NEXT:    v_mov_b32_e32 v54, v0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; kill: killed $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr0
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(33)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    ; implicit-def: $vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB90_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v16
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v16
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v15
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v15
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v14
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v14
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v14
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v13
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v13
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v12
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v12
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v12
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v11
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v11
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v10
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[9:10]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v10
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v10
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v9
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[7:8]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v9
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v8
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v8
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v8
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v7
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v6
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v6
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v4
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v4
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v2
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v58
; GFX9-NEXT:    s_waitcnt vmcnt(62)
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[57:58]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v58
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v58
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v57
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v57
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v30
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v30
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v29
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v28
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v28
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v27
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v26
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v26
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v23
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v23
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v20
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v31, 8, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 8, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v40, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v41, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v26
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 16, v25
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v25
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 24, v24
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v24
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 24, v22
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v22
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 24, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 8, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 24, v18
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v17
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 8, v17
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:  .LBB90_2: ; %Flow
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    v_mov_b32_e32 v34, v52
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    s_xor_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB90_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    v_lshlrev_b32_e32 v31, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT:    v_bfe_u32 v32, v31, 16, 1
; GFX9-NEXT:    s_movk_i32 s6, 0x7fff
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_add3_u32 v32, v32, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_bfe_u32 v31, v18, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v34, v32, v33, vcc
; GFX9-NEXT:    v_add3_u32 v31, v31, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v32, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v48, v31, v32, vcc
; GFX9-NEXT:    v_bfe_u32 v31, v18, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_add3_u32 v31, v31, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v32, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v49, v31, v32, vcc
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v31, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_mov_b32 s7, 0x7060302
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v31, vcc
; GFX9-NEXT:    v_perm_b32 v32, v17, v49, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v20
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_perm_b32 v33, v48, v34, s7
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v31, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v20
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v18, v31, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v19
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v38, v18, v20, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v19
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v37, v18, v20, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_perm_b32 v31, v17, v37, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v22
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v22
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v59, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v21
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v21
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v60, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_perm_b32 v15, v17, v60, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v24
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v24
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_perm_b32 v16, v0, v59, s7
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v23
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v23
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_perm_b32 v20, v15, v0, s7
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_perm_b32 v19, v17, v0, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v26
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_mov_b32_e32 v36, v20
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_mov_b32_e32 v35, v19
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v26
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v25
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v25
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_perm_b32 v16, v15, v0, s7
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_perm_b32 v15, v17, v0, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v28
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v28
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v26, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v27
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v28, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v27
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v25, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_perm_b32 v15, v17, v25, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v30
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v30
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v27, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v29
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v30, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v29
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v24, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_perm_b32 v16, v28, v26, s7
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_perm_b32 v15, v17, v24, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v58
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v58
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v29, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_waitcnt vmcnt(48)
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v57
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v58, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v57
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v23, v18, v19, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_perm_b32 v16, v30, v27, s7
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_perm_b32 v15, v17, v23, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v17, 16, v2
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_add3_u32 v18, v18, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    v_bfe_u32 v17, v2, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v57, v18, v19, vcc
; GFX9-NEXT:    v_add3_u32 v17, v17, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v21, v17, v18, vcc
; GFX9-NEXT:    v_bfe_u32 v17, v2, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_add3_u32 v17, v17, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v22, v17, v18, vcc
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_perm_b32 v16, v58, v29, s7
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v17, vcc
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_perm_b32 v61, v1, v22, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v4
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v17, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v4
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v2, v17, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v2, v4, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v3
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v2, v4, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_perm_b32 v39, v1, v19, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v6
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v2, v3, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v2, v3, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v5
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v4, v2, v3, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_perm_b32 v41, v1, v4, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v8
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v8
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v2, v3, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v7
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v8, v2, v3, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v7
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v2, v3, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v7, vcc
; GFX9-NEXT:    v_perm_b32 v43, v1, v3, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v10
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v7, v2, v7, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v10, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v9
; GFX9-NEXT:    v_perm_b32 v32, v38, v33, s7
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_mov_b32_e32 v51, v32
; GFX9-NEXT:    v_cndmask_b32_e32 v10, v2, v10, vcc
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_mov_b32_e32 v50, v31
; GFX9-NEXT:    v_add3_u32 v2, v2, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v31, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v9
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_bfe_u32 v9, v1, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v2, v31, vcc
; GFX9-NEXT:    v_add3_u32 v9, v9, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v31, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v9, v31, vcc
; GFX9-NEXT:    v_perm_b32 v45, v1, v2, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v12
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_bfe_u32 v9, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v9, v9, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v31, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v12
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_bfe_u32 v12, v1, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v9, v9, v31, vcc
; GFX9-NEXT:    v_add3_u32 v12, v12, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v31, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 16, v11
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v12, v12, v31, vcc
; GFX9-NEXT:    v_bfe_u32 v31, v1, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT:    v_add3_u32 v31, v31, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v32, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v31, v32, vcc
; GFX9-NEXT:    v_bfe_u32 v31, v11, 16, 1
; GFX9-NEXT:    v_add3_u32 v31, v31, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v32, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v31, v32, vcc
; GFX9-NEXT:    v_perm_b32 v55, v11, v1, s7
; GFX9-NEXT:    v_lshlrev_b32_e32 v11, 16, v14
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_bfe_u32 v31, v11, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT:    v_add3_u32 v31, v31, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v32, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v31, v32, vcc
; GFX9-NEXT:    v_bfe_u32 v31, v14, 16, 1
; GFX9-NEXT:    v_add3_u32 v31, v31, v14, s6
; GFX9-NEXT:    v_or_b32_e32 v32, 0x400000, v14
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v14, v31, v32, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v31, 16, v13
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT:    v_bfe_u32 v52, v31, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_add3_u32 v52, v52, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v53, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_bfe_u32 v31, v13, 16, 1
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v52, v53, vcc
; GFX9-NEXT:    v_add3_u32 v31, v31, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v52, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v31, v52, vcc
; GFX9-NEXT:    v_perm_b32 v31, v13, v0, s7
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v16
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_bfe_u32 v52, v13, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT:    v_add3_u32 v52, v52, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v53, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v52, v53, vcc
; GFX9-NEXT:    v_bfe_u32 v52, v16, 16, 1
; GFX9-NEXT:    v_add3_u32 v52, v52, v16, s6
; GFX9-NEXT:    v_or_b32_e32 v53, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v52, v53, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v52, 16, v15
; GFX9-NEXT:    v_add_f32_e32 v52, 0x40c00000, v52
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT:    v_bfe_u32 v47, v52, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX9-NEXT:    v_add3_u32 v47, v47, v52, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v52
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX9-NEXT:    v_bfe_u32 v52, v15, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v47, v47, v63, vcc
; GFX9-NEXT:    v_add3_u32 v52, v52, v15, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v15
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v52, v63, vcc
; GFX9-NEXT:    v_perm_b32 v53, v16, v13, s7
; GFX9-NEXT:    v_perm_b32 v52, v15, v47, s7
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v16
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 16, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v13, 16, v14
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v13, 16, v0
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v12
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v10
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v8
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v6
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v21
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v58
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v30
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v28
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshrrev_b32_e32 v28, 16, v26
; GFX9-NEXT:    v_lshrrev_b32_e32 v30, 16, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v27, 16, v25
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v47
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 16, v29
; GFX9-NEXT:    v_lshrrev_b32_e32 v29, 16, v24
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v23
; GFX9-NEXT:    v_perm_b32 v62, v21, v57, s7
; GFX9-NEXT:    v_perm_b32 v56, v12, v9, s7
; GFX9-NEXT:    v_lshrrev_b32_e32 v12, 16, v9
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v57
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v40, v18, v20, s7
; GFX9-NEXT:    v_perm_b32 v42, v6, v17, s7
; GFX9-NEXT:    v_perm_b32 v46, v10, v7, s7
; GFX9-NEXT:    v_lshrrev_b32_e32 v10, 16, v7
; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v17
; GFX9-NEXT:    v_lshrrev_b32_e32 v3, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 16, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v49
; GFX9-NEXT:    v_perm_b32 v44, v8, v5, s7
; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v4, 16, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 16, v33
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v34
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    v_perm_b32 v32, v14, v11, s7
; GFX9-NEXT:    v_lshrrev_b32_e32 v14, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v11, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v22
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v59
; GFX9-NEXT:    v_lshrrev_b32_e32 v21, 16, v60
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[52:53]
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(11)
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshrrev_b32_e32 v26, 16, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[55:56]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[45:46]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[43:44]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[41:42]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[57:58]
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v38
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v48
; GFX9-NEXT:    v_lshrrev_b64 v[48:49], 24, v[39:40]
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[48:49], 24, v[61:62]
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[37:38]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[33:34]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[47:48]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[35:36]
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v53
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v53
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v52
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v52
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v32
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v32
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v31
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v31
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v56
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v56
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v55
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v55
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v46
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v46
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v45
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v45
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v44
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v44
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v43
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v42
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v40
; GFX9-NEXT:    v_mov_b32_e32 v45, v57
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v62
; GFX9-NEXT:    v_mov_b32_e32 v46, v58
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v46
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v46
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v45
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v45
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v38
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v38
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v37
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v37
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshrrev_b32_e32 v31, 8, v43
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 8, v42
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 16, v61
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v61
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v34
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v34
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v33
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v33
; GFX9-NEXT:    v_mov_b32_e32 v33, v47
; GFX9-NEXT:    v_mov_b32_e32 v34, v48
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(39)
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[35:36]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[50:51]
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 8, v62
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 24, v34
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 16, v41
; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 8, v41
; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 16, v39
; GFX9-NEXT:    v_lshrrev_b32_e32 v41, 8, v39
; GFX9-NEXT:    v_mov_b32_e32 v58, v16
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 16, v33
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v33
; GFX9-NEXT:    v_lshrrev_b32_e32 v40, 8, v40
; GFX9-NEXT:    v_mov_b32_e32 v57, v15
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 24, v36
; GFX9-NEXT:    s_waitcnt vmcnt(15)
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v37
; GFX9-NEXT:    s_waitcnt vmcnt(14)
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v38
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 8, v38
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v37
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v35
; GFX9-NEXT:    s_waitcnt vmcnt(8)
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v49
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_lshrrev_b64 v[59:60], 24, v[49:50]
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v36
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v35
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 24, v62
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 8, v62
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 16, v61
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v61
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 24, v50
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 8, v50
; GFX9-NEXT:    v_mov_b32_e32 v50, v39
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 8, v49
; GFX9-NEXT:    v_mov_b32_e32 v49, v33
; GFX9-NEXT:  .LBB90_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v7, v7, v31, s4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v3, v3, v41, s4
; GFX9-NEXT:    v_perm_b32 v4, v4, v40, s4
; GFX9-NEXT:    v_perm_b32 v5, v5, v32, s4
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v2, v2, v51, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v1, v1, v44, s4
; GFX9-NEXT:    v_perm_b32 v6, v6, v43, s4
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v17, v17, v39, s4
; GFX9-NEXT:    v_perm_b32 v0, v18, v0, s4
; GFX9-NEXT:    s_waitcnt vmcnt(8)
; GFX9-NEXT:    v_perm_b32 v9, v9, v33, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(8)
; GFX9-NEXT:    v_perm_b32 v8, v8, v31, s4
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_perm_b32 v32, v32, v40, s4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_perm_b32 v51, v53, v51, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_perm_b32 v42, v42, v61, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v42, 16, v1
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_perm_b32 v10, v10, v33, s4
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v31, v31, v40, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v53, v56, v52, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v52, v33, v40, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v11, v33, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v12, v12, v33, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v40, v33, v40, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v33, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v14, v33, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v41, v33, v43, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v15, v15, v33, s4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v43, v63, v43, s4
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v16, v16, v33, s4
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v18, v45, v62, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v33, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v2
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshl_or_b32 v1, v51, 16, v3
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshl_or_b32 v1, v53, 16, v5
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v6
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshl_or_b32 v1, v32, 16, v7
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v8
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshl_or_b32 v1, v31, 16, v9
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v10
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshl_or_b32 v1, v52, 16, v11
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v12
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshl_or_b32 v1, v40, 16, v13
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v14
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshl_or_b32 v1, v41, 16, v15
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v16
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshl_or_b32 v1, v43, 16, v17
; GFX9-NEXT:    buffer_store_dword v1, v54, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v35, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:68
; GFX9-NEXT:    v_perm_b32 v0, v19, v46, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v18, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v20, v37, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v36, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v21, v59, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v38, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v22, v60, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v48, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v23, v0, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v24, v50, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v49, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v25, v34, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v55, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v26, v47, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v27, v0, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v0, v28, v0, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v29, v0, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v0, v30, v0, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v57, v0, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v0, v58, v0, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    buffer_store_dword v0, v54, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v128i8:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x2
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v81, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v80, off, s32
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v89, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v90, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v91, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v92, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v93, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v94, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v95, s32 offset:112
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v104, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v105, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v106, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v107, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v108, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v109, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v110, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v111, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v120, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v121, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v122, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v123, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v124, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v125, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v126, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v127, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v136, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v137, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v138, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v139, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v140, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v141, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v142, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v143, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v152, s32 offset:12
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr141_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr139_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr138_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr136_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr122_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr121_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr120_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr111_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr109_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr104_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr95_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr93_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr94_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr90_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr88_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr127_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr105_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr142_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr137_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr152_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr143_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr140_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr126_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr125_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr124_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr123_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr110_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr108_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr107_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr106_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr92_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr91_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr89_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr31_hi16
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB90_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[99:100], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[116:117], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[100:101], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[117:118], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[130:131], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[133:134], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[96:97], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[101:102], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v56, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v72, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v78, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v88, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v90, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v93, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v95, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v104, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v109, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v111, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v120, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v121, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v122, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v136, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v138, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v139, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v141, 8, v1
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v81
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v81
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v80
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v44, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v60, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v75, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v77, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v79, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v89, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v91, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v92, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v106, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v107, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v108, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v110, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v123, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v124, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v125, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v126, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v140, 8, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[86:87], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[112:113], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[134:135], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[82:83], 24, v[80:81]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[97:98], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[102:103], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[118:119], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[131:132], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.h, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v64.h, v1.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.h, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v65.h, v2.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v180.h, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v66.h, v3.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v179.h, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v67.h, v4.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.h, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v68.h, v5.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.h, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v69.h, v6.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.h, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.h, v7.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.h, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.h, v8.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v94.h, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.h, v9.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v63.h, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.h, v10.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v127.h, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.h, v11.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v105.h, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v115.h, v12.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v142.h, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.h, v13.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v137.h, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.h, v14.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v152.h, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v144.h, v15.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v143.h, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.h, v16.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.h, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v17.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.h, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.h, v18.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v148.h, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.h, v19.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.h, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.h, v20.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v150.h, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.h, v21.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.h, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v22.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.h, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v23.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v151.h, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.h, v24.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.h, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v48.h, v25.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v161.h, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v49.h, v26.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.h, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v27.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v163.h, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v28.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.h, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v52.h, v29.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.h, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.h, v30.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v176.h, v80.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v54.h, v80.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v167.h, v81.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.h, v81.h
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr1
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr3
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr5
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr7
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr9
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr11
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr13
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr15
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr19
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr21
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr23
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr25
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr27
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr29
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-TRUE16-NEXT:  .LBB90_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB90_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v34, 0x40c00000, v32 :: v_dual_lshlrev_b32 v31, 16, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v31, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v31
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v31, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v31, v38, v34, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v39, v32, v37, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_cndmask_b32 v32, v35, v36
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v39.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v48, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v125, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v126, 8, v32
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v146, v37, v49, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v33
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v31, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v34 :: v_dual_add_f32 v19, 0x40c00000, v19
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v33, v33, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v31.l, v146.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v147, v33, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_add3_u32 v20, v20, v36, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v17, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v34.l, v147.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v140, 8, v31
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v19, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v148, v17, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v33, v20, v35 :: v_dual_and_b32 v20, 0xffff0000, v21
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v110, 24, v34
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v123, 8, v34
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_lshlrev_b32 v22, 16, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v149, v19, v35 :: v_dual_lshlrev_b32 v22, 16, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v21
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v17, v36, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v36.l, v149.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v21, 0x7fff
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v150, v17, v24 :: v_dual_lshlrev_b32 v21, 16, v23
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v35, v19, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_add_f32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v106, 24, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v107, 8, v36
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v151, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v17, v24, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v38.l, v151.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v33.l, v148.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v160, v17, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v37, v19, v24, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v124, 8, v33
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v161, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v28
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v22, 0x40c00000, v22 :: v_dual_cndmask_b32 v49, v17, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v35.l, v150.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v49.l, v161.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v162, v17, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_cndmask_b32 v48, v19, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v75, 24, v49
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v77, 8, v49
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v108, 8, v35
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v163, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v30
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v22, 0x40c00000, v22 :: v_dual_cndmask_b32 v51, v17, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v29
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v51.l, v163.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v89, 24, v38
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v91, 8, v38
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v164, v17, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_cndmask_b32 v50, v19, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v22, 0x7fff
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v81
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v60, 24, v51
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v51
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v165, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v81
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v22, 0x40c00000, v22 :: v_dual_cndmask_b32 v53, v17, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v21, 0x7fff
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v80
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v53.l, v165.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v166, v17, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v80
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_cndmask_b32 v52, v19, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v37.l, v160.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v44, 24, v53
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v47, 8, v53
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v167, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v55, v17, v24, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v55.l, v167.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v92, 8, v37
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v176, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v19, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v55
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v42, 8, v55
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v20, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v177, v19, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v65, v17, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v17, v17, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v65.l, v177.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v178, v17, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v64, v2, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v48.l, v162.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v138, 24, v65
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v179, v2, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v67, v1, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v67.l, v179.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v50.l, v164.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v121, 24, v67
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v180, v1, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v66, v2, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v122, 8, v67
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v139, 8, v65
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v50
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v79, 8, v48
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v181, v2, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v69, v1, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v69.l, v181.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v182, v2, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v5, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v109, 24, v69
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v68, v1, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v2, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v68.l, v182.h
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_cndmask_b32 v45, v3, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v71, v1, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v4, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v52.l, v166.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v71.l, v45.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v70, v1, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v1, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v62, v4, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v70.l, v62.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v63, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v85, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v9, 16, v13
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v85.l, v63.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v54.l, v176.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[96:97], 24, v[52:53]
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v84, v2, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v7, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[82:83], 24, v[54:55]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[97:98], 24, v[48:49]
; GFX11-TRUE16-NEXT:    v_add3_u32 v1, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v94, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v7, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v84.l, v94.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v105, v1, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[116:117], 24, v[84:85]
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v115, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v115.l, v105.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[117:118], 24, v[70:71]
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v114, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v4 :: v_dual_add_f32 v3, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v7, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[118:119], 24, v[33:34]
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v127, v4, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v7, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v6, v8, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v129, v4, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_add3_u32 v3, v8, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v16
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v137, v6, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v5, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v129.l, v137.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v128, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v16
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v6, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v66.l, v180.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v64.l, v178.h
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v142, v4, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v4, v7, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v2, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v7, v7, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, v142.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v143, v2, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, v127.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[99:100], 24, v[128:129]
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v152, v7, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_add3_u32 v2, v10, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[100:101], 24, v[50:51]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[133:134], 24, v[66:67]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[101:102], 24, v[37:38]
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v145, v4, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v145.l, v143.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[130:131], 24, v[68:69]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[112:113], 24, v[114:115]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[134:135], 24, v[64:65]
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v144, v2, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v144.l, v152.h
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[102:103], 24, v[35:36]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[131:132], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v145
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v145
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[86:87], 24, v[144:145]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v46, 8, v144
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v56, 24, v129
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v129
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v128
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v72, 24, v115
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v115
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v114
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v78, 24, v85
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v88, 8, v85
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v90, 8, v84
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v93, 24, v71
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v95, 8, v71
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v104, 8, v70
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v111, 8, v69
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v120, 8, v68
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v136, 8, v66
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v141, 8, v64
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v54
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v57, 8, v52
; GFX11-TRUE16-NEXT:  .LBB90_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v1.l, v178.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v64.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v3.l, v177.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v65.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v5.l, v180.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v141, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v3, v139, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v138, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v67.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v7.l, v182.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v68.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v4, 16, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v5, v136, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v66.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v5.l, v179.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v121, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v120, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v122, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v9.l, v181.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v69.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v71.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v6, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v8, 16, v7
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v9, v111, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v10, v109, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v62.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v70.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v45.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v84.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v8, v104, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v9, v117, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v10, v95, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v11, v93, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v94.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v63.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v8, 16, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v114.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v10, 16, v9
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v11, v90, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v12, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v13, v88, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v85.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v127.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v105.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v115.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v78, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v10, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v5.l, v142.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v12, 16, v11
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v14, 16, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v6.l, v128.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v7.l, v137.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v8.l, v129.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v9.l, v152.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v10.l, v144.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v11.l, v143.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v12.l, v145.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v13.l, v146.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v31.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v74, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v72, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v59, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v99, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v58, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v56, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v41, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v183, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v140, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v16, 16, v15
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v39.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v32.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v12.l, v148.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v33.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v14.l, v147.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v34.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v16.l, v150.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v35.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v18.l, v149.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v36.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v126, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v125, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v124, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v123, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v110, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v108, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v107, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v106, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v15.l, v160.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v37.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v17.l, v151.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v38.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v19.l, v162.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v48.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v21.l, v161.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v49.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v23.l, v164.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v50.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v92, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v91, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v89, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v79, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v77, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v75, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v20.l, v163.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v51.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v22.l, v166.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v52.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v24.l, v165.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v53.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v26.l, v176.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v54.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v28.l, v167.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v55.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v60, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v22, v57, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v23, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v44, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v26, v43, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v28, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v29, v40, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-TRUE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v152, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_b32 v143, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_b32 v142, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v141, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_b32 v140, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v139, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v138, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_b32 v137, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_b32 v136, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_b32 v127, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_b32 v126, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_b32 v125, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_b32 v124, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_b32 v123, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_b32 v122, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_b32 v121, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_b32 v120, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_b32 v111, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_b32 v110, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_b32 v109, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_b32 v108, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_b32 v107, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_b32 v106, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_b32 v105, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_b32 v104, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_b32 v95, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_b32 v94, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_b32 v93, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_b32 v92, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_b32 v91, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_b32 v90, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_b32 v89, off, s32 offset:136
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:236
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v128i8:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x2
; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_clause 0x16 ; 92-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:12
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB90_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v1
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 8, v32
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v77, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v78, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[17:18]
; GFX11-FAKE16-NEXT:  .LBB90_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB90_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v34, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v34, 0x40c00000, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v35, 0x40c00000, v17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v17, v17, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v38, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v17, v17, v36 :: v_dual_add_f32 v18, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v48, v35, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v18, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v34
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v58, v37, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v68, v58, v17, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v38, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 24, v68
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v68
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v36, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v67, v34, v33, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v77, 16, v67
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v78, 8, v67
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v18, 0x7fff
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v35, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v38, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v34, v20, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v20, v39, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v38
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v34, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_perm_b32 v70, v19, v18, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v19
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v20, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v20, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v69, v35, v34, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v37, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v35, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v36, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v22, v48, v37, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v24
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v39, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v39
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v35, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v37
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v39, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 24, v70
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 8, v70
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v35, v22, v35 :: v_dual_add_f32 v22, 0x40c00000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v81, v21, v20, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v23
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v22, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v35
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v69
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v69
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v37, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v48, 16, 1
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v80, v36, v35, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v48, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[69:70]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 24, v81
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v80
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v36, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v24, v49, v38, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v48
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v36, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v23
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v24, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v39
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v39, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v83, v23, v22, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v37, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v38, v24, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v82, v37, v36, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v49, 0x40c00000, v25
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v39, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v38, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v26, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v49, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v26, v50, v39, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v49
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v37, v25, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v39
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v38, v49, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 24, v83
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v37, v26, v37 :: v_dual_add_f32 v26, 0x40c00000, v48
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v26, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v38, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v26
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-FAKE16-NEXT:    v_add3_u32 v39, v39, v26, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v84, v38, v37, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v28, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v50, 0x40c00000, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v28
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v39, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v38, v28, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v50, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v28, v51, v48, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v49, 16, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v50
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v38, v27, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v48
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_add3_u32 v39, v39, v50, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v36
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v27
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v28, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v49
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v30, 0x40c00000, v30 :: v_dual_lshlrev_b32 v49, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v28, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v28
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v28, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v86, v39, v38, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v51, 0x40c00000, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v49, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v48, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v39, v39, v30, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v51, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v30, v52, v49, 0x7fff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v51
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v39, v29, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v49
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v51, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v32
; GFX11-FAKE16-NEXT:    v_perm_b32 v85, v25, v24, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v37
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v39, v30, v39 :: v_dual_add_f32 v30, 0x40c00000, v50
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v48, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v50, 0x40c00000, v50
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v49, v30, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v96, v48, v39, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v52, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v50, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v49, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v48, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v32, v53, v50, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v51, 16, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v48, v31, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v50
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v49, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v87, v27, v26, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v31
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v32, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v32, 0x40c00000, v51
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v51, 16, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v49, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v51, 0x40c00000, v51
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v50, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v98, v49, v48, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v53, 0x40c00000, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v51, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v50, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v49, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v32, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v51
; GFX11-FAKE16-NEXT:    v_perm_b32 v97, v29, v28, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v38
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v49, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v55, v51, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v32, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v49, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v49, 0x40c00000, v54
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v39
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 24, v97
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v49, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v52, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v52, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v49
; GFX11-FAKE16-NEXT:    v_add3_u32 v51, v51, v49, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v100, v50, v32, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v52, 0x40c00000, v52
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v50, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v51, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 16, v100
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v50, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v55, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    v_add3_u32 v51, v51, v3, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v52, 0x40c00000, v54
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v50, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v53, 16, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v51, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v53, 0x40c00000, v53
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v102, v3, v50, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v51, v51, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v3, v3, v6, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v51, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add3_u32 v6, v64, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v52, v5, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v55, 16, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v99, v31, v30, 0x7060302
; GFX11-FAKE16-NEXT:    v_perm_b32 v113, v3, v51, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v48
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v53, v6, v54 :: v_dual_add_f32 v6, 0x40c00000, v55
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v52, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v54, 0x40c00000, v54
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v112, v5, v53, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v52, v6, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_bfe_u32 v65, v54, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v101, v2, v1, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v5, v5, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v52, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v8, v65, v54, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v54
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v5, v64 :: v_dual_lshlrev_b32 v64, 16, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v52, v7, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v8, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v8, 0x40c00000, v64
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v55, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v52, v65, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v55, 0x40c00000, v55
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v179, v7, v54, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v52, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v66, v55, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v103, v4, v49, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v7, v7, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v52, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v10, v66, v55, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v55
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v65, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v52, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v30
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v55, v10, v64 :: v_dual_add_f32 v10, 0x40c00000, v65
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v26
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v52, v66 :: v_dual_lshlrev_b32 v52, 16, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v65, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v176, v9, v55, 0x7060302
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v52
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v64, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_add3_u32 v65, v65, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v9, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v11, 0x40c00000, v11 :: v_dual_cndmask_b32 v10, v52, v64
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v71, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v177, v7, v8, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v65, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v65, 16, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v66, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v18
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v52, v64 :: v_dual_add_f32 v52, 0x40c00000, v65
; GFX11-FAKE16-NEXT:    v_add3_u32 v64, v66, v11, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v180, v5, v6, 0x7060302
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v64, v65 :: v_dual_add_f32 v64, 0x40c00000, v66
; GFX11-FAKE16-NEXT:    v_add3_u32 v65, v71, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    v_bfe_u32 v114, v64, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v115, 0x400000, v64
; GFX11-FAKE16-NEXT:    v_perm_b32 v162, v11, v9, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v65, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v65, v71, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_add3_u32 v71, v114, v64, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v114, 16, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v116, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v33
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v65, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v65, 0x40c00000, v114
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-FAKE16-NEXT:    v_add3_u32 v66, v116, v13, 0x7fff
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v150, v14, v52, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v114, v65, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v71, v115, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v115, 16, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v163, v12, v10, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v49
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v66, v71, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v66, 0x40c00000, v115
; GFX11-FAKE16-NEXT:    v_add3_u32 v71, v114, v65, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v114, 0x400000, v65
; GFX11-FAKE16-NEXT:    v_bfe_u32 v115, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v116, v66, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v117, 0x400000, v66
; GFX11-FAKE16-NEXT:    v_perm_b32 v149, v13, v64, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v71, v114, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v114, v115, v16, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v115, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v116, v116, v66, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v118, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v64
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v114, v115, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_add3_u32 v71, v71, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v52
; GFX11-FAKE16-NEXT:    v_perm_b32 v135, v16, v65, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v116, v117, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v53
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[112:113]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[102:103]
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v71, v118, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v65
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[100:101]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v134, v15, v66, 0x7060302
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v66
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v54
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v50
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[134:135]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[149:150]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[162:163]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[176:177]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[179:180]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[96:97]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[84:85]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[80:81]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v55
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v51
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[98:99]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[86:87]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[82:83]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[67:68]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v135
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 8, v135
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v134
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v134
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v150
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v150
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v149
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v149
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 24, v163
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v163
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 16, v162
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 8, v162
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 24, v177
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v177
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v176
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v176
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v180
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 8, v180
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v179
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v179
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v113
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v113
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v112
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v112
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 24, v103
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v103
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v102
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v102
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 24, v101
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 8, v101
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 24, v99
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 8, v99
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v98
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v98
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v97
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 16, v96
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 8, v96
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 24, v87
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v87
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v86
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v86
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 24, v85
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v85
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v84
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v84
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v83
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v82
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v82
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v81
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 8, v80
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v58
; GFX11-FAKE16-NEXT:  .LBB90_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v43, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v51, v129, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v183, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v119, v182, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v177, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v51, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v48, v181, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v51, v118, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v117, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v37, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v164, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v116, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v151, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v115, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v48, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v51, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v52, 16, v8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v36, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v37, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v35, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v13, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v146, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v14, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v114, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v15, v134, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v133, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v16, v132, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v71, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v17, v78, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v77, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v18, v76, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v68, v75, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v19, v74, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v73, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v20, v72, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v178, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v21, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v62, v65, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v166, v61, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v23, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v60, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v24, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v161, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v25, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v57, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v26, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v148, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v47, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v28, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v144, v46, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v29, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v45, v49, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v30, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v130, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v31, v98, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v101, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v128, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x16 ; 92-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:100
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define inreg <128 x i8> @bitcast_v64bf16_to_v128i8_scalar(<64 x bfloat> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64bf16_to_v128i8_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_writelane_b32 v63, s70, 22
; SI-NEXT:    v_writelane_b32 v63, s71, 23
; SI-NEXT:    v_writelane_b32 v63, s80, 24
; SI-NEXT:    v_writelane_b32 v63, s81, 25
; SI-NEXT:    v_writelane_b32 v63, s82, 26
; SI-NEXT:    v_writelane_b32 v63, s83, 27
; SI-NEXT:    v_writelane_b32 v63, s84, 28
; SI-NEXT:    v_writelane_b32 v63, s85, 29
; SI-NEXT:    v_readfirstlane_b32 s56, v15
; SI-NEXT:    v_writelane_b32 v63, s86, 30
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    v_writelane_b32 v63, s87, 31
; SI-NEXT:    v_readfirstlane_b32 s46, v16
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    v_readfirstlane_b32 s78, v8
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v3
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s57
; SI-NEXT:    v_writelane_b32 v63, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s44, v17
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_and_b32 s30, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s56
; SI-NEXT:    v_writelane_b32 v63, s97, 33
; SI-NEXT:    v_readfirstlane_b32 s42, v18
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    v_readfirstlane_b32 s58, v14
; SI-NEXT:    v_readfirstlane_b32 s60, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v12
; SI-NEXT:    v_readfirstlane_b32 s72, v11
; SI-NEXT:    v_readfirstlane_b32 s74, v10
; SI-NEXT:    v_readfirstlane_b32 s76, v9
; SI-NEXT:    v_readfirstlane_b32 s88, v7
; SI-NEXT:    v_readfirstlane_b32 s90, v6
; SI-NEXT:    v_readfirstlane_b32 s92, v5
; SI-NEXT:    v_readfirstlane_b32 s94, v4
; SI-NEXT:    s_and_b32 s34, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v1
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s46
; SI-NEXT:    v_writelane_b32 v63, s98, 34
; SI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s29, 16
; SI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s7, s28, 16
; SI-NEXT:    s_and_b32 s8, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s9, s27, 16
; SI-NEXT:    s_and_b32 s10, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s11, s26, 16
; SI-NEXT:    s_and_b32 s12, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s13, s25, 16
; SI-NEXT:    s_and_b32 s14, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s15, s24, 16
; SI-NEXT:    s_and_b32 s24, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s25, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s26, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s27, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s28, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s29, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_and_b32 s41, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_and_b32 s63, s62, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_and_b32 s73, s72, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    s_and_b32 s75, s74, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_and_b32 s77, s76, 0xffff0000
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_and_b32 s79, s78, 0xffff0000
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_and_b32 s89, s88, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_and_b32 s91, s90, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_and_b32 s93, s92, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_and_b32 s95, s94, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_and_b32 s36, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v19
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s44
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    v_writelane_b32 v63, s99, 35
; SI-NEXT:    s_cmp_lg_u32 vcc_lo, 0
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s41
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s16
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s40
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s17
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s29
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s18
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s28
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s19
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s27
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s20
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s26
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s21
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s25
; SI-NEXT:    v_mul_f32_e64 v52, 1.0, s22
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s24
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s23
; SI-NEXT:    v_mul_f32_e64 v34, 1.0, s14
; SI-NEXT:    v_mul_f32_e64 v45, 1.0, s15
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s12
; SI-NEXT:    v_mul_f32_e64 v33, 1.0, s13
; SI-NEXT:    v_mul_f32_e64 v35, 1.0, s10
; SI-NEXT:    v_mul_f32_e64 v49, 1.0, s11
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_mul_f32_e64 v56, 1.0, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v60, 1.0, s9
; SI-NEXT:    v_mul_f32_e64 v54, 1.0, s6
; SI-NEXT:    v_mul_f32_e64 v40, 1.0, s7
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s4
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s5
; SI-NEXT:    v_mul_f32_e64 v38, 1.0, s36
; SI-NEXT:    v_mul_f32_e64 v42, 1.0, s37
; SI-NEXT:    v_mul_f32_e64 v48, 1.0, s34
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s35
; SI-NEXT:    v_mul_f32_e64 v51, 1.0, s30
; SI-NEXT:    v_mul_f32_e64 v50, 1.0, s31
; SI-NEXT:    v_mul_f32_e64 v53, 1.0, s95
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s94
; SI-NEXT:    v_mul_f32_e64 v44, 1.0, s93
; SI-NEXT:    v_mul_f32_e64 v47, 1.0, s92
; SI-NEXT:    v_mul_f32_e64 v41, 1.0, s91
; SI-NEXT:    v_mul_f32_e64 v32, 1.0, s90
; SI-NEXT:    v_mul_f32_e64 v57, 1.0, s89
; SI-NEXT:    v_mul_f32_e64 v59, 1.0, s88
; SI-NEXT:    v_mul_f32_e64 v46, 1.0, s79
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s78
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s77
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s76
; SI-NEXT:    v_mul_f32_e64 v58, 1.0, s75
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s74
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s73
; SI-NEXT:    v_mul_f32_e64 v36, 1.0, s72
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s63
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s62
; SI-NEXT:    v_mul_f32_e64 v43, 1.0, s61
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s60
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s59
; SI-NEXT:    v_mul_f32_e64 v55, 1.0, s58
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s47
; SI-NEXT:    v_mul_f32_e64 v39, 1.0, s45
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s43
; SI-NEXT:    v_mul_f32_e64 v37, 1.0, s42
; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; SI-NEXT:    ; implicit-def: $vgpr61 : SGPR spill to VGPR lane
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB91_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_readfirstlane_b32 s4, v10
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v15
; SI-NEXT:    s_lshr_b64 s[8:9], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v7
; SI-NEXT:    s_lshr_b32 s7, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v24
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v25
; SI-NEXT:    s_lshr_b64 s[86:87], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v9
; SI-NEXT:    s_lshr_b32 s65, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v27
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v29
; SI-NEXT:    v_mov_b32_e32 v10, v21
; SI-NEXT:    s_lshr_b64 s[80:81], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v10
; SI-NEXT:    s_lshr_b32 s69, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v31
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v52
; SI-NEXT:    s_lshr_b64 s[66:67], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v14
; SI-NEXT:    s_lshr_b32 s91, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v34
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v45
; SI-NEXT:    s_lshr_b64 s[52:53], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v20
; SI-NEXT:    s_lshr_b32 s37, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v35
; SI-NEXT:    v_mov_b32_e32 v35, v49
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v35
; SI-NEXT:    s_lshr_b64 s[30:31], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v56
; SI-NEXT:    v_mov_b32_e32 v52, v54
; SI-NEXT:    s_lshr_b32 s89, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v52
; SI-NEXT:    v_mov_b32_e32 v45, v40
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v45
; SI-NEXT:    v_mov_b32_e32 v21, v22
; SI-NEXT:    v_readfirstlane_b32 s6, v4
; SI-NEXT:    s_lshr_b64 s[50:51], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v21
; SI-NEXT:    v_mov_b32_e32 v4, v38
; SI-NEXT:    s_lshr_b32 s57, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v4
; SI-NEXT:    v_mov_b32_e32 v54, v42
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v54
; SI-NEXT:    v_mov_b32_e32 v22, v48
; SI-NEXT:    s_lshr_b64 s[92:93], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v22
; SI-NEXT:    v_mov_b32_e32 v40, v51
; SI-NEXT:    s_lshr_b32 s79, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v40
; SI-NEXT:    v_mov_b32_e32 v42, v50
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v42
; SI-NEXT:    v_mov_b32_e32 v49, v2
; SI-NEXT:    v_mov_b32_e32 v2, v12
; SI-NEXT:    s_lshr_b64 s[76:77], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v53
; SI-NEXT:    v_mov_b32_e32 v12, v44
; SI-NEXT:    s_lshr_b32 s73, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v12
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v47
; SI-NEXT:    s_lshr_b64 s[62:63], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v41
; SI-NEXT:    s_lshr_b32 s59, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v57
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v59
; SI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v46
; SI-NEXT:    s_lshr_b32 s45, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v3
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v1
; SI-NEXT:    v_mov_b32_e32 v44, v32
; SI-NEXT:    s_lshr_b64 s[40:41], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v58
; SI-NEXT:    v_mov_b32_e32 v32, v13
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b32 s29, s4, 16
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_mov_b32_e32 v1, v9
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_mov_b32_e32 v9, v6
; SI-NEXT:    v_readfirstlane_b32 s4, v6
; SI-NEXT:    v_mov_b32_e32 v6, v36
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v6
; SI-NEXT:    v_mov_b32_e32 v3, v23
; SI-NEXT:    s_lshr_b64 s[24:25], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v5
; SI-NEXT:    v_mov_b32_e32 v23, v43
; SI-NEXT:    s_lshr_b32 s23, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v23
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v32
; SI-NEXT:    s_lshr_b64 s[16:17], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v11
; SI-NEXT:    s_lshr_b32 s19, s4, 16
; SI-NEXT:    s_lshr_b64 s[96:97], s[6:7], 16
; SI-NEXT:    s_mov_b32 s9, s96
; SI-NEXT:    v_readfirstlane_b32 s88, v60
; SI-NEXT:    s_lshr_b64 s[82:83], s[88:89], 16
; SI-NEXT:    v_readfirstlane_b32 s64, v19
; SI-NEXT:    s_lshr_b64 s[84:85], s[64:65], 16
; SI-NEXT:    s_mov_b32 s87, s84
; SI-NEXT:    v_readfirstlane_b32 s68, v26
; SI-NEXT:    s_lshr_b64 s[70:71], s[68:69], 16
; SI-NEXT:    s_mov_b32 s81, s70
; SI-NEXT:    v_readfirstlane_b32 s90, v30
; SI-NEXT:    s_lshr_b64 s[38:39], s[90:91], 16
; SI-NEXT:    s_mov_b32 s67, s38
; SI-NEXT:    v_readfirstlane_b32 s36, v33
; SI-NEXT:    s_lshr_b64 s[98:99], s[36:37], 16
; SI-NEXT:    s_mov_b32 s53, s98
; SI-NEXT:    s_mov_b32 s31, s82
; SI-NEXT:    v_readfirstlane_b32 s56, v49
; SI-NEXT:    s_lshr_b64 s[94:95], s[56:57], 16
; SI-NEXT:    s_mov_b32 s51, s94
; SI-NEXT:    v_readfirstlane_b32 s78, v2
; SI-NEXT:    s_lshr_b64 s[74:75], s[78:79], 16
; SI-NEXT:    s_mov_b32 s93, s74
; SI-NEXT:    v_mov_b32_e32 v38, v16
; SI-NEXT:    v_readfirstlane_b32 s72, v38
; SI-NEXT:    s_lshr_b64 s[60:61], s[72:73], 16
; SI-NEXT:    s_mov_b32 s77, s60
; SI-NEXT:    v_readfirstlane_b32 s58, v44
; SI-NEXT:    s_lshr_b64 s[54:55], s[58:59], 16
; SI-NEXT:    s_mov_b32 s63, s54
; SI-NEXT:    v_mov_b32_e32 v15, v18
; SI-NEXT:    v_readfirstlane_b32 s44, v15
; SI-NEXT:    s_lshr_b64 s[42:43], s[44:45], 16
; SI-NEXT:    s_mov_b32 s47, s42
; SI-NEXT:    v_readfirstlane_b32 s28, v3
; SI-NEXT:    s_lshr_b64 s[26:27], s[28:29], 16
; SI-NEXT:    s_mov_b32 s41, s26
; SI-NEXT:    v_mov_b32_e32 v18, v28
; SI-NEXT:    v_readfirstlane_b32 s22, v18
; SI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 16
; SI-NEXT:    s_mov_b32 s25, s34
; SI-NEXT:    v_mov_b32_e32 v28, v55
; SI-NEXT:    v_readfirstlane_b32 s18, v28
; SI-NEXT:    s_lshr_b64 s[14:15], s[18:19], 16
; SI-NEXT:    s_mov_b32 s17, s14
; SI-NEXT:    v_lshrrev_b32_e32 v25, 24, v7
; SI-NEXT:    v_lshrrev_b32_e32 v24, 24, v1
; SI-NEXT:    v_lshrrev_b32_e32 v7, 24, v20
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v53
; SI-NEXT:    v_lshrrev_b32_e32 v26, 24, v8
; SI-NEXT:    s_lshr_b32 s78, s96, 8
; SI-NEXT:    s_lshr_b32 s61, s84, 8
; SI-NEXT:    s_lshr_b32 s72, s70, 8
; SI-NEXT:    s_lshr_b32 s75, s38, 8
; SI-NEXT:    s_lshr_b32 s58, s98, 8
; SI-NEXT:    s_lshr_b32 s43, s82, 8
; SI-NEXT:    s_lshr_b32 s44, s94, 8
; SI-NEXT:    s_mov_b32 s64, s74
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_readfirstlane_b32 s4, v13
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_readfirstlane_b32 s12, v48
; SI-NEXT:    s_lshr_b32 s27, s74, 8
; SI-NEXT:    s_mov_b32 s90, s60
; SI-NEXT:    s_lshr_b32 s28, s60, 8
; SI-NEXT:    s_lshr_b32 s74, s54, 8
; SI-NEXT:    s_mov_b32 s68, s42
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s4, v36
; SI-NEXT:    s_lshr_b64 s[10:11], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s4, v17
; SI-NEXT:    s_lshr_b32 s13, s4, 16
; SI-NEXT:    s_mov_b32 s5, s13
; SI-NEXT:    v_writelane_b32 v61, s4, 26
; SI-NEXT:    v_writelane_b32 v61, s5, 27
; SI-NEXT:    v_readfirstlane_b32 s4, v39
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v51
; SI-NEXT:    s_lshr_b64 s[4:5], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s5, v8
; SI-NEXT:    s_lshr_b64 s[20:21], s[12:13], 16
; SI-NEXT:    s_lshr_b32 s13, s5, 16
; SI-NEXT:    v_readfirstlane_b32 s12, v37
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 16
; SI-NEXT:    s_mov_b32 s5, vcc_lo
; SI-NEXT:    s_mov_b32 s88, vcc_lo
; SI-NEXT:    s_lshr_b32 s6, vcc_lo, 8
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 4
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 5
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 2
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 3
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 0
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 1
; SI-NEXT:    s_lshr_b64 vcc, s[86:87], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 10
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 11
; SI-NEXT:    s_lshr_b64 vcc, s[86:87], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 8
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 9
; SI-NEXT:    s_lshr_b64 vcc, s[86:87], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 6
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 7
; SI-NEXT:    s_lshr_b64 vcc, s[80:81], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 16
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 17
; SI-NEXT:    s_lshr_b64 vcc, s[80:81], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 14
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 15
; SI-NEXT:    s_lshr_b64 vcc, s[80:81], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 12
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 13
; SI-NEXT:    s_lshr_b64 vcc, s[66:67], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 22
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 23
; SI-NEXT:    s_lshr_b64 vcc, s[66:67], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 20
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 21
; SI-NEXT:    s_lshr_b64 vcc, s[66:67], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 18
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 19
; SI-NEXT:    s_lshr_b64 vcc, s[52:53], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 28
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 29
; SI-NEXT:    s_lshr_b64 vcc, s[52:53], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 26
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 27
; SI-NEXT:    s_lshr_b64 vcc, s[52:53], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 24
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 25
; SI-NEXT:    s_lshr_b64 vcc, s[30:31], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 34
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 35
; SI-NEXT:    s_lshr_b64 vcc, s[30:31], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 32
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 33
; SI-NEXT:    s_lshr_b64 vcc, s[30:31], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 30
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 31
; SI-NEXT:    s_lshr_b64 vcc, s[50:51], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 40
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 41
; SI-NEXT:    s_lshr_b64 vcc, s[50:51], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 38
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 39
; SI-NEXT:    s_lshr_b64 vcc, s[50:51], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 36
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 37
; SI-NEXT:    s_lshr_b64 vcc, s[92:93], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 46
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 47
; SI-NEXT:    s_lshr_b64 vcc, s[92:93], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 44
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 45
; SI-NEXT:    s_lshr_b64 vcc, s[92:93], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 42
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 43
; SI-NEXT:    s_lshr_b64 vcc, s[76:77], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 52
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 53
; SI-NEXT:    s_lshr_b64 vcc, s[76:77], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 50
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 51
; SI-NEXT:    s_lshr_b64 vcc, s[76:77], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 48
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 49
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 24
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 58
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 59
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 56
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 57
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 54
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 55
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 24
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 0
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 1
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 16
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 62
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 63
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 8
; SI-NEXT:    v_writelane_b32 v62, vcc_lo, 60
; SI-NEXT:    v_writelane_b32 v62, vcc_hi, 61
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 24
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 6
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 7
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 16
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 4
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 5
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 8
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 2
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 3
; SI-NEXT:    s_lshr_b64 vcc, s[24:25], 24
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 12
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 13
; SI-NEXT:    s_lshr_b64 vcc, s[24:25], 16
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 10
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 11
; SI-NEXT:    s_lshr_b64 vcc, s[24:25], 8
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 8
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 9
; SI-NEXT:    s_lshr_b64 vcc, s[16:17], 24
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 18
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 19
; SI-NEXT:    s_lshr_b64 vcc, s[16:17], 16
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 16
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 17
; SI-NEXT:    s_lshr_b64 vcc, s[16:17], 8
; SI-NEXT:    s_mov_b32 s11, s20
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 14
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 15
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 24
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 24
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 25
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 16
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 22
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 23
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 8
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 20
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 21
; SI-NEXT:    s_lshr_b64 vcc, s[4:5], 24
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 32
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 33
; SI-NEXT:    s_lshr_b64 vcc, s[4:5], 16
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 30
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 31
; SI-NEXT:    s_lshr_b64 vcc, s[4:5], 8
; SI-NEXT:    v_writelane_b32 v61, vcc_lo, 28
; SI-NEXT:    s_lshr_b32 s22, s42, 8
; SI-NEXT:    s_mov_b32 s56, s26
; SI-NEXT:    s_lshr_b32 s21, s26, 8
; SI-NEXT:    s_lshr_b32 s18, s34, 8
; SI-NEXT:    s_mov_b32 s36, s14
; SI-NEXT:    s_lshr_b32 s15, s14, 8
; SI-NEXT:    s_mov_b32 s14, s20
; SI-NEXT:    s_lshr_b32 s12, s20, 8
; SI-NEXT:    v_writelane_b32 v61, vcc_hi, 29
; SI-NEXT:    s_mov_b64 vcc, 0
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v10
; SI-NEXT:    v_lshrrev_b32_e32 v43, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v55, 24, v56
; SI-NEXT:    v_lshrrev_b32_e32 v50, 24, v21
; SI-NEXT:    v_lshrrev_b32_e32 v60, 24, v22
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v34, 24, v41
; SI-NEXT:    v_lshrrev_b32_e32 v33, 24, v46
; SI-NEXT:    v_lshrrev_b32_e32 v31, 24, v58
; SI-NEXT:    v_lshrrev_b32_e32 v30, 24, v5
; SI-NEXT:    v_lshrrev_b32_e32 v29, 24, v11
; SI-NEXT:    v_lshrrev_b32_e32 v27, 24, v17
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_mov_b32_e32 v20, v35
; SI-NEXT:    v_mov_b32_e32 v14, v52
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_mov_b32_e32 v37, v45
; SI-NEXT:    v_mov_b32_e32 v45, v2
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v39, v4
; SI-NEXT:    v_mov_b32_e32 v52, v44
; SI-NEXT:    v_mov_b32_e32 v44, v12
; SI-NEXT:    v_mov_b32_e32 v35, v3
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, v18
; SI-NEXT:    v_mov_b32_e32 v12, v6
; SI-NEXT:    v_mov_b32_e32 v6, v10
; SI-NEXT:    v_mov_b32_e32 v10, v7
; SI-NEXT:    v_mov_b32_e32 v18, v28
; SI-NEXT:    v_mov_b32_e32 v16, v23
; SI-NEXT:    v_mov_b32_e32 v23, v32
; SI-NEXT:    v_mov_b32_e32 v28, v48
; SI-NEXT:    v_mov_b32_e32 v32, v13
; SI-NEXT:    v_mov_b32_e32 v48, v22
; SI-NEXT:    v_mov_b32_e32 v7, v5
; SI-NEXT:    v_mov_b32_e32 v2, v11
; SI-NEXT:    v_mov_b32_e32 v3, v17
; SI-NEXT:    s_branch .LBB91_3
; SI-NEXT:  .LBB91_2:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v37, v40
; SI-NEXT:    v_writelane_b32 v62, s4, 0
; SI-NEXT:    v_writelane_b32 v62, s5, 1
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v45, v12
; SI-NEXT:    v_writelane_b32 v62, s4, 2
; SI-NEXT:    v_writelane_b32 v62, s5, 3
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v40, v51
; SI-NEXT:    v_writelane_b32 v62, s4, 4
; SI-NEXT:    v_writelane_b32 v62, s5, 5
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v52, v32
; SI-NEXT:    v_writelane_b32 v62, s4, 6
; SI-NEXT:    v_writelane_b32 v62, s5, 7
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v1, v28
; SI-NEXT:    v_writelane_b32 v62, s4, 8
; SI-NEXT:    v_writelane_b32 v62, s5, 9
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v12, v36
; SI-NEXT:    v_writelane_b32 v62, s4, 10
; SI-NEXT:    v_writelane_b32 v62, s5, 11
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_writelane_b32 v62, s4, 12
; SI-NEXT:    v_writelane_b32 v62, s5, 13
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_writelane_b32 v62, s4, 14
; SI-NEXT:    v_writelane_b32 v62, s5, 15
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr21
; SI-NEXT:    ; implicit-def: $sgpr20
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:    v_mov_b32_e32 v9, v6
; SI-NEXT:    v_writelane_b32 v62, s4, 16
; SI-NEXT:    v_writelane_b32 v62, s5, 17
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    s_mov_b64 vcc, -1
; SI-NEXT:    v_writelane_b32 v62, s4, 18
; SI-NEXT:    v_writelane_b32 v62, s5, 19
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v20, v49
; SI-NEXT:    v_writelane_b32 v62, s4, 20
; SI-NEXT:    v_writelane_b32 v62, s5, 21
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v49, v2
; SI-NEXT:    v_writelane_b32 v62, s4, 22
; SI-NEXT:    v_writelane_b32 v62, s5, 23
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v14, v54
; SI-NEXT:    v_writelane_b32 v62, s4, 24
; SI-NEXT:    v_writelane_b32 v62, s5, 25
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v39, v38
; SI-NEXT:    v_writelane_b32 v62, s4, 26
; SI-NEXT:    v_writelane_b32 v62, s5, 27
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v54, v42
; SI-NEXT:    v_writelane_b32 v62, s4, 28
; SI-NEXT:    v_writelane_b32 v62, s5, 29
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v38, v16
; SI-NEXT:    v_writelane_b32 v62, s4, 30
; SI-NEXT:    v_writelane_b32 v62, s5, 31
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v42, v50
; SI-NEXT:    v_writelane_b32 v62, s4, 32
; SI-NEXT:    v_writelane_b32 v62, s5, 33
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v15, v18
; SI-NEXT:    v_writelane_b32 v62, s4, 34
; SI-NEXT:    v_writelane_b32 v62, s5, 35
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v35, v23
; SI-NEXT:    v_writelane_b32 v62, s4, 36
; SI-NEXT:    v_writelane_b32 v62, s5, 37
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v6, v21
; SI-NEXT:    v_writelane_b32 v62, s4, 38
; SI-NEXT:    v_writelane_b32 v62, s5, 39
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v18, v55
; SI-NEXT:    v_writelane_b32 v62, s4, 40
; SI-NEXT:    v_writelane_b32 v62, s5, 41
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v16, v43
; SI-NEXT:    v_writelane_b32 v62, s4, 42
; SI-NEXT:    v_writelane_b32 v62, s5, 43
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v23, v13
; SI-NEXT:    v_writelane_b32 v62, s4, 44
; SI-NEXT:    v_writelane_b32 v62, s5, 45
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v21, v22
; SI-NEXT:    v_writelane_b32 v62, s4, 46
; SI-NEXT:    v_writelane_b32 v62, s5, 47
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v7, v5
; SI-NEXT:    v_writelane_b32 v62, s4, 48
; SI-NEXT:    v_writelane_b32 v62, s5, 49
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v2, v11
; SI-NEXT:    v_writelane_b32 v62, s4, 50
; SI-NEXT:    v_writelane_b32 v62, s5, 51
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_mov_b32_e32 v3, v17
; SI-NEXT:    v_writelane_b32 v62, s4, 52
; SI-NEXT:    v_writelane_b32 v62, s5, 53
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr7
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr61
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr75
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr37
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr43
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr27
; SI-NEXT:    ; implicit-def: $sgpr28
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $sgpr22
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $sgpr18
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $sgpr15
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; kill: killed $vgpr26
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr57
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr73
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr59
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr45
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr29
; SI-NEXT:    ; implicit-def: $sgpr24
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr23
; SI-NEXT:    ; implicit-def: $sgpr16
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr19
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr13
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    v_writelane_b32 v62, s4, 54
; SI-NEXT:    v_writelane_b32 v62, s5, 55
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v62, s4, 56
; SI-NEXT:    v_writelane_b32 v62, s5, 57
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v62, s4, 58
; SI-NEXT:    v_writelane_b32 v62, s5, 59
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v62, s4, 60
; SI-NEXT:    v_writelane_b32 v62, s5, 61
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v62, s4, 62
; SI-NEXT:    v_writelane_b32 v62, s5, 63
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 0
; SI-NEXT:    v_writelane_b32 v61, s5, 1
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 2
; SI-NEXT:    v_writelane_b32 v61, s5, 3
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 4
; SI-NEXT:    v_writelane_b32 v61, s5, 5
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 6
; SI-NEXT:    v_writelane_b32 v61, s5, 7
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 8
; SI-NEXT:    v_writelane_b32 v61, s5, 9
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 10
; SI-NEXT:    v_writelane_b32 v61, s5, 11
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 12
; SI-NEXT:    v_writelane_b32 v61, s5, 13
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 14
; SI-NEXT:    v_writelane_b32 v61, s5, 15
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 16
; SI-NEXT:    v_writelane_b32 v61, s5, 17
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 18
; SI-NEXT:    v_writelane_b32 v61, s5, 19
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 20
; SI-NEXT:    v_writelane_b32 v61, s5, 21
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 22
; SI-NEXT:    v_writelane_b32 v61, s5, 23
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s4, 24
; SI-NEXT:    v_writelane_b32 v61, s5, 25
; SI-NEXT:    ; implicit-def: $sgpr5
; SI-NEXT:    v_writelane_b32 v61, s4, 26
; SI-NEXT:    v_writelane_b32 v61, s5, 27
; SI-NEXT:    v_writelane_b32 v61, s20, 28
; SI-NEXT:    v_writelane_b32 v61, s21, 29
; SI-NEXT:    ; implicit-def: $sgpr20
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v61, s20, 30
; SI-NEXT:    v_writelane_b32 v61, s21, 31
; SI-NEXT:    v_writelane_b32 v61, s88, 32
; SI-NEXT:    v_writelane_b32 v61, s89, 33
; SI-NEXT:    ; implicit-def: $sgpr88
; SI-NEXT:  .LBB91_3: ; %Flow
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, vcc
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v26, v27
; SI-NEXT:    v_mov_b32_e32 v27, v29
; SI-NEXT:    v_mov_b32_e32 v29, v30
; SI-NEXT:    v_mov_b32_e32 v30, v31
; SI-NEXT:    v_mov_b32_e32 v31, v33
; SI-NEXT:    s_cbranch_vccnz .LBB91_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_readfirstlane_b32 s4, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v8
; SI-NEXT:    s_lshr_b32 s5, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v5
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_readfirstlane_b32 s12, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v36
; SI-NEXT:    v_readfirstlane_b32 s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v10
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_readfirstlane_b32 s6, v11
; SI-NEXT:    s_lshr_b64 s[10:11], s[8:9], 16
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v28
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_mov_b32 s7, s9
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_writelane_b32 v61, s6, 26
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v23
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v2
; SI-NEXT:    v_writelane_b32 v61, s7, 27
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_readfirstlane_b32 s6, v8
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v9
; SI-NEXT:    s_lshr_b64 s[20:21], s[8:9], 16
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v5
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b32 s19, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_readfirstlane_b32 s18, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    s_lshr_b64 s[16:17], s[8:9], 16
; SI-NEXT:    v_readfirstlane_b32 s8, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v7
; SI-NEXT:    v_readfirstlane_b32 s22, v3
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    s_lshr_b64 s[24:25], s[8:9], 16
; SI-NEXT:    v_readfirstlane_b32 s6, v2
; SI-NEXT:    s_lshr_b32 s23, s6, 16
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v57
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v44
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v39
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 16
; SI-NEXT:    s_mov_b32 s25, s34
; SI-NEXT:    s_lshr_b64 s[26:27], s[18:19], 16
; SI-NEXT:    s_mov_b32 s17, s26
; SI-NEXT:    s_mov_b32 s11, s20
; SI-NEXT:    s_lshr_b64 s[4:5], s[4:5], 16
; SI-NEXT:    v_readfirstlane_b32 s5, v17
; SI-NEXT:    s_lshr_b32 s13, s5, 16
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 16
; SI-NEXT:    s_mov_b32 s5, vcc_lo
; SI-NEXT:    s_lshr_b32 s18, s34, 8
; SI-NEXT:    s_lshr_b32 s12, s20, 8
; SI-NEXT:    v_lshrrev_b32_e32 v19, 24, v12
; SI-NEXT:    v_lshrrev_b32_e32 v29, 24, v2
; SI-NEXT:    v_lshrrev_b32_e32 v27, 24, v5
; SI-NEXT:    v_lshrrev_b32_e32 v26, 24, v11
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_readfirstlane_b32 s8, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v35
; SI-NEXT:    v_readfirstlane_b32 s6, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v3
; SI-NEXT:    v_readfirstlane_b32 s28, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v59
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v58
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b64 s[40:41], s[8:9], 16
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b32 s29, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v7
; SI-NEXT:    v_readfirstlane_b32 s44, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v47
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v46
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b64 s[46:47], s[8:9], 16
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v52
; SI-NEXT:    v_readfirstlane_b32 s6, v7
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b32 s45, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v8
; SI-NEXT:    v_readfirstlane_b32 s58, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v42
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b64 s[62:63], s[8:9], 16
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v38
; SI-NEXT:    v_readfirstlane_b32 s6, v8
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b32 s59, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v9
; SI-NEXT:    v_readfirstlane_b32 s72, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v54
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v53
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b64 s[76:77], s[8:9], 16
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v45
; SI-NEXT:    v_readfirstlane_b32 s6, v9
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b32 s73, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v10
; SI-NEXT:    v_readfirstlane_b32 s78, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v37
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b64 s[92:93], s[8:9], 16
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v14
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v49
; SI-NEXT:    v_readfirstlane_b32 s6, v13
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    s_lshr_b32 s79, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v21
; SI-NEXT:    v_readfirstlane_b32 s56, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v20
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v10
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b64 s[50:51], s[8:9], 16
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    v_readfirstlane_b32 s6, v14
; SI-NEXT:    s_lshr_b32 s57, s6, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b64 s[94:95], s[56:57], 16
; SI-NEXT:    s_mov_b32 s51, s94
; SI-NEXT:    s_lshr_b64 s[48:49], s[78:79], 16
; SI-NEXT:    s_mov_b32 s93, s48
; SI-NEXT:    s_lshr_b64 s[74:75], s[72:73], 16
; SI-NEXT:    s_mov_b32 s77, s74
; SI-NEXT:    s_lshr_b64 s[54:55], s[58:59], 16
; SI-NEXT:    s_mov_b32 s63, s54
; SI-NEXT:    s_lshr_b64 s[60:61], s[44:45], 16
; SI-NEXT:    s_mov_b32 s47, s60
; SI-NEXT:    s_lshr_b64 s[42:43], s[28:29], 16
; SI-NEXT:    s_mov_b32 s41, s42
; SI-NEXT:    s_lshr_b32 s44, s94, 8
; SI-NEXT:    s_lshr_b32 s27, s48, 8
; SI-NEXT:    s_lshr_b32 s28, s74, 8
; SI-NEXT:    s_lshr_b32 s22, s60, 8
; SI-NEXT:    s_mov_b32 s56, s42
; SI-NEXT:    s_lshr_b32 s21, s42, 8
; SI-NEXT:    v_lshrrev_b32_e32 v50, 24, v14
; SI-NEXT:    v_lshrrev_b32_e32 v60, 24, v13
; SI-NEXT:    v_lshrrev_b32_e32 v34, 24, v8
; SI-NEXT:    v_lshrrev_b32_e32 v31, 24, v7
; SI-NEXT:    v_lshrrev_b32_e32 v30, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_readfirstlane_b32 s6, v10
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v4
; SI-NEXT:    v_readfirstlane_b32 s88, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[8:9], 16
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v56
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_readfirstlane_b32 s6, v4
; SI-NEXT:    s_lshr_b32 s89, s6, 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[88:89], 16
; SI-NEXT:    s_mov_b32 s31, s82
; SI-NEXT:    s_lshr_b32 s43, s82, 8
; SI-NEXT:    s_mov_b32 s88, vcc_lo
; SI-NEXT:    v_lshrrev_b32_e32 v55, 24, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_readfirstlane_b32 s8, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v15
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    s_lshr_b64 s[52:53], s[8:9], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v10
; SI-NEXT:    v_readfirstlane_b32 s36, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_readfirstlane_b32 s6, v10
; SI-NEXT:    s_lshr_b32 s37, s6, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b64 s[98:99], s[36:37], 16
; SI-NEXT:    s_mov_b32 s53, s98
; SI-NEXT:    s_lshr_b32 s58, s98, 8
; SI-NEXT:    s_mov_b32 s36, s26
; SI-NEXT:    v_lshrrev_b32_e32 v10, 24, v10
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_readfirstlane_b32 s8, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_readfirstlane_b32 s6, v16
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    s_lshr_b64 s[66:67], s[8:9], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v15
; SI-NEXT:    v_readfirstlane_b32 s90, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_readfirstlane_b32 s6, v15
; SI-NEXT:    s_lshr_b32 s91, s6, 16
; SI-NEXT:    s_lshr_b64 s[38:39], s[90:91], 16
; SI-NEXT:    s_mov_b32 s67, s38
; SI-NEXT:    s_lshr_b32 s75, s38, 8
; SI-NEXT:    s_mov_b32 s90, s74
; SI-NEXT:    s_lshr_b32 s74, s54, 8
; SI-NEXT:    v_lshrrev_b32_e32 v43, 24, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v24, 24, v6
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_readfirstlane_b32 s68, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    s_lshr_b64 s[80:81], s[8:9], 16
; SI-NEXT:    v_readfirstlane_b32 s6, v12
; SI-NEXT:    s_lshr_b32 s69, s6, 16
; SI-NEXT:    s_lshr_b64 s[70:71], s[68:69], 16
; SI-NEXT:    s_mov_b32 s81, s70
; SI-NEXT:    s_lshr_b32 s72, s70, 8
; SI-NEXT:    s_mov_b32 s68, s60
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v25, 24, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_readfirstlane_b32 s64, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    s_lshr_b64 s[86:87], s[8:9], 16
; SI-NEXT:    v_readfirstlane_b32 s6, v6
; SI-NEXT:    s_lshr_b32 s65, s6, 16
; SI-NEXT:    s_lshr_b64 s[84:85], s[64:65], 16
; SI-NEXT:    s_mov_b32 s87, s84
; SI-NEXT:    s_lshr_b32 s61, s84, 8
; SI-NEXT:    s_mov_b32 s64, s48
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_readfirstlane_b32 s8, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    v_readfirstlane_b32 s6, v18
; SI-NEXT:    s_lshr_b32 s9, s6, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v1
; SI-NEXT:    s_lshr_b32 s7, s6, 16
; SI-NEXT:    s_lshr_b64 s[8:9], s[8:9], 16
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_readfirstlane_b32 s6, v16
; SI-NEXT:    s_lshr_b64 s[96:97], s[6:7], 16
; SI-NEXT:    s_mov_b32 s9, s96
; SI-NEXT:    s_lshr_b64 s[14:15], s[8:9], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 4
; SI-NEXT:    v_writelane_b32 v62, s15, 5
; SI-NEXT:    s_lshr_b64 s[14:15], s[8:9], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 2
; SI-NEXT:    v_writelane_b32 v62, s15, 3
; SI-NEXT:    s_lshr_b64 s[14:15], s[8:9], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 0
; SI-NEXT:    v_writelane_b32 v62, s15, 1
; SI-NEXT:    s_lshr_b64 s[14:15], s[86:87], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 10
; SI-NEXT:    v_writelane_b32 v62, s15, 11
; SI-NEXT:    s_lshr_b64 s[14:15], s[86:87], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 8
; SI-NEXT:    v_writelane_b32 v62, s15, 9
; SI-NEXT:    s_lshr_b64 s[14:15], s[86:87], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 6
; SI-NEXT:    v_writelane_b32 v62, s15, 7
; SI-NEXT:    s_lshr_b64 s[14:15], s[80:81], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 16
; SI-NEXT:    v_writelane_b32 v62, s15, 17
; SI-NEXT:    s_lshr_b64 s[14:15], s[80:81], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 14
; SI-NEXT:    v_writelane_b32 v62, s15, 15
; SI-NEXT:    s_lshr_b64 s[14:15], s[80:81], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 12
; SI-NEXT:    v_writelane_b32 v62, s15, 13
; SI-NEXT:    s_lshr_b64 s[14:15], s[66:67], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 22
; SI-NEXT:    v_writelane_b32 v62, s15, 23
; SI-NEXT:    s_lshr_b64 s[14:15], s[66:67], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 20
; SI-NEXT:    v_writelane_b32 v62, s15, 21
; SI-NEXT:    s_lshr_b64 s[14:15], s[66:67], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 18
; SI-NEXT:    v_writelane_b32 v62, s15, 19
; SI-NEXT:    s_lshr_b64 s[14:15], s[52:53], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 28
; SI-NEXT:    v_writelane_b32 v62, s15, 29
; SI-NEXT:    s_lshr_b64 s[14:15], s[52:53], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 26
; SI-NEXT:    v_writelane_b32 v62, s15, 27
; SI-NEXT:    s_lshr_b64 s[14:15], s[52:53], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 24
; SI-NEXT:    v_writelane_b32 v62, s15, 25
; SI-NEXT:    s_lshr_b64 s[14:15], s[30:31], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 34
; SI-NEXT:    v_writelane_b32 v62, s15, 35
; SI-NEXT:    s_lshr_b64 s[14:15], s[30:31], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 32
; SI-NEXT:    v_writelane_b32 v62, s15, 33
; SI-NEXT:    s_lshr_b64 s[14:15], s[30:31], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 30
; SI-NEXT:    v_writelane_b32 v62, s15, 31
; SI-NEXT:    s_lshr_b64 s[14:15], s[50:51], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 40
; SI-NEXT:    v_writelane_b32 v62, s15, 41
; SI-NEXT:    s_lshr_b64 s[14:15], s[50:51], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 38
; SI-NEXT:    v_writelane_b32 v62, s15, 39
; SI-NEXT:    s_lshr_b64 s[14:15], s[50:51], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 36
; SI-NEXT:    v_writelane_b32 v62, s15, 37
; SI-NEXT:    s_lshr_b64 s[14:15], s[92:93], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 46
; SI-NEXT:    v_writelane_b32 v62, s15, 47
; SI-NEXT:    s_lshr_b64 s[14:15], s[92:93], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 44
; SI-NEXT:    v_writelane_b32 v62, s15, 45
; SI-NEXT:    s_lshr_b64 s[14:15], s[92:93], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 42
; SI-NEXT:    v_writelane_b32 v62, s15, 43
; SI-NEXT:    s_lshr_b64 s[14:15], s[76:77], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 52
; SI-NEXT:    v_writelane_b32 v62, s15, 53
; SI-NEXT:    s_lshr_b64 s[14:15], s[76:77], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 50
; SI-NEXT:    v_writelane_b32 v62, s15, 51
; SI-NEXT:    s_lshr_b64 s[14:15], s[76:77], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 48
; SI-NEXT:    v_writelane_b32 v62, s15, 49
; SI-NEXT:    s_lshr_b64 s[14:15], s[62:63], 24
; SI-NEXT:    v_writelane_b32 v62, s14, 58
; SI-NEXT:    v_writelane_b32 v62, s15, 59
; SI-NEXT:    s_lshr_b64 s[14:15], s[62:63], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 56
; SI-NEXT:    v_writelane_b32 v62, s15, 57
; SI-NEXT:    s_lshr_b64 s[14:15], s[62:63], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 54
; SI-NEXT:    v_writelane_b32 v62, s15, 55
; SI-NEXT:    s_lshr_b64 s[14:15], s[46:47], 24
; SI-NEXT:    v_writelane_b32 v61, s14, 0
; SI-NEXT:    v_writelane_b32 v61, s15, 1
; SI-NEXT:    s_lshr_b64 s[14:15], s[46:47], 16
; SI-NEXT:    v_writelane_b32 v62, s14, 62
; SI-NEXT:    v_writelane_b32 v62, s15, 63
; SI-NEXT:    s_lshr_b64 s[14:15], s[46:47], 8
; SI-NEXT:    v_writelane_b32 v62, s14, 60
; SI-NEXT:    v_writelane_b32 v62, s15, 61
; SI-NEXT:    s_lshr_b64 s[14:15], s[40:41], 24
; SI-NEXT:    v_writelane_b32 v61, s14, 6
; SI-NEXT:    v_writelane_b32 v61, s15, 7
; SI-NEXT:    s_lshr_b64 s[14:15], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v61, s14, 4
; SI-NEXT:    v_writelane_b32 v61, s15, 5
; SI-NEXT:    s_lshr_b64 s[14:15], s[40:41], 8
; SI-NEXT:    v_writelane_b32 v61, s14, 2
; SI-NEXT:    v_writelane_b32 v61, s15, 3
; SI-NEXT:    s_lshr_b64 s[14:15], s[24:25], 24
; SI-NEXT:    v_writelane_b32 v61, s14, 12
; SI-NEXT:    v_writelane_b32 v61, s15, 13
; SI-NEXT:    s_lshr_b64 s[14:15], s[24:25], 16
; SI-NEXT:    v_writelane_b32 v61, s14, 10
; SI-NEXT:    v_writelane_b32 v61, s15, 11
; SI-NEXT:    s_lshr_b64 s[14:15], s[24:25], 8
; SI-NEXT:    v_writelane_b32 v61, s14, 8
; SI-NEXT:    v_writelane_b32 v61, s15, 9
; SI-NEXT:    s_lshr_b64 s[14:15], s[16:17], 24
; SI-NEXT:    v_writelane_b32 v61, s14, 18
; SI-NEXT:    v_writelane_b32 v61, s15, 19
; SI-NEXT:    s_lshr_b64 s[14:15], s[16:17], 16
; SI-NEXT:    v_writelane_b32 v61, s14, 16
; SI-NEXT:    v_writelane_b32 v61, s15, 17
; SI-NEXT:    s_lshr_b64 s[14:15], s[16:17], 8
; SI-NEXT:    v_writelane_b32 v61, s14, 14
; SI-NEXT:    v_writelane_b32 v61, s15, 15
; SI-NEXT:    s_lshr_b64 s[14:15], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v61, s14, 24
; SI-NEXT:    v_writelane_b32 v61, s15, 25
; SI-NEXT:    s_lshr_b64 s[14:15], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v61, s14, 22
; SI-NEXT:    v_writelane_b32 v61, s15, 23
; SI-NEXT:    s_lshr_b64 s[14:15], s[10:11], 8
; SI-NEXT:    v_writelane_b32 v61, s14, 20
; SI-NEXT:    v_writelane_b32 v61, s15, 21
; SI-NEXT:    s_lshr_b64 s[14:15], s[4:5], 24
; SI-NEXT:    v_writelane_b32 v61, s14, 32
; SI-NEXT:    v_writelane_b32 v61, s15, 33
; SI-NEXT:    s_lshr_b64 s[14:15], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v61, s14, 30
; SI-NEXT:    v_writelane_b32 v61, s15, 31
; SI-NEXT:    s_lshr_b64 s[14:15], s[4:5], 8
; SI-NEXT:    v_writelane_b32 v61, s14, 28
; SI-NEXT:    v_writelane_b32 v61, s15, 29
; SI-NEXT:    s_lshr_b32 s78, s96, 8
; SI-NEXT:    s_lshr_b32 s15, s26, 8
; SI-NEXT:    s_mov_b32 s14, s20
; SI-NEXT:    s_lshr_b32 s6, vcc_lo, 8
; SI-NEXT:  .LBB91_5: ; %end
; SI-NEXT:    s_and_b32 s5, s8, 0xff
; SI-NEXT:    v_readlane_b32 s8, v62, 0
; SI-NEXT:    v_readlane_b32 s9, v62, 1
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 2
; SI-NEXT:    v_readlane_b32 s9, v62, 3
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 vcc_lo, v62, 4
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, vcc_lo, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, s5
; SI-NEXT:    s_and_b32 s5, s96, 0xff
; SI-NEXT:    s_lshl_b32 s8, s78, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s7, 0xff
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v25
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 6
; SI-NEXT:    v_readlane_b32 vcc_hi, v62, 5
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s86, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 7
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 8
; SI-NEXT:    v_readlane_b32 s9, v62, 9
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 vcc_lo, v62, 10
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, vcc_lo, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_readlane_b32 vcc_hi, v62, 11
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s84, 0xff
; SI-NEXT:    s_lshl_b32 s8, s61, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s65, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 12
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s80, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 13
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 14
; SI-NEXT:    v_readlane_b32 s9, v62, 15
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s60, v62, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s60, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s70, 0xff
; SI-NEXT:    s_lshl_b32 s8, s72, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 16, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s69, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v19
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 18
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s66, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 19
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    v_readlane_b32 s61, v62, 17
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 20
; SI-NEXT:    v_readlane_b32 s9, v62, 21
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s60, v62, 22
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s60, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s38, 0xff
; SI-NEXT:    s_lshl_b32 s8, s75, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 24, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s91, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v43
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 24
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s52, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 25
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    v_readlane_b32 s61, v62, 23
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 26
; SI-NEXT:    v_readlane_b32 s9, v62, 27
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s60, v62, 28
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s60, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s98, 0xff
; SI-NEXT:    s_lshl_b32 s8, s58, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s37, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v10
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 30
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s30, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 31
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    v_readlane_b32 s61, v62, 29
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 32
; SI-NEXT:    v_readlane_b32 s9, v62, 33
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s60, v62, 34
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s60, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s82, 0xff
; SI-NEXT:    s_lshl_b32 s8, s43, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 40, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s89, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v55
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 36
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s50, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 37
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 38
; SI-NEXT:    v_readlane_b32 s9, v62, 39
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s42, v62, 40
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s42, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s94, 0xff
; SI-NEXT:    s_lshl_b32 s8, s44, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 48, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s57, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v50
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 42
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s92, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 43
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    v_readlane_b32 s43, v62, 41
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 44
; SI-NEXT:    v_readlane_b32 s9, v62, 45
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s42, v62, 46
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s42, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s64, 0xff
; SI-NEXT:    s_lshl_b32 s8, s27, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 56, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s79, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v60
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 48
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s76, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 49
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 50
; SI-NEXT:    v_readlane_b32 s9, v62, 51
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s26, v62, 52
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s26, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 64, v0
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s5, s90, 0xff
; SI-NEXT:    s_lshl_b32 s8, s28, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s73, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_readlane_b32 s27, v62, 53
; SI-NEXT:    v_readlane_b32 s26, v62, 58
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    v_readlane_b32 s27, v62, 59
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    v_readlane_b32 s61, v62, 35
; SI-NEXT:    v_readlane_b32 s43, v62, 47
; SI-NEXT:    v_readlane_b32 s99, v63, 35
; SI-NEXT:    v_readlane_b32 s98, v63, 34
; SI-NEXT:    v_readlane_b32 s97, v63, 33
; SI-NEXT:    v_readlane_b32 s96, v63, 32
; SI-NEXT:    v_readlane_b32 s87, v63, 31
; SI-NEXT:    v_readlane_b32 s86, v63, 30
; SI-NEXT:    v_readlane_b32 s85, v63, 29
; SI-NEXT:    v_readlane_b32 s84, v63, 28
; SI-NEXT:    v_readlane_b32 s83, v63, 27
; SI-NEXT:    v_readlane_b32 s82, v63, 26
; SI-NEXT:    v_readlane_b32 s81, v63, 25
; SI-NEXT:    v_readlane_b32 s80, v63, 24
; SI-NEXT:    v_readlane_b32 s71, v63, 23
; SI-NEXT:    v_readlane_b32 s70, v63, 22
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 54
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s62, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 55
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 56
; SI-NEXT:    v_readlane_b32 s9, v62, 57
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s26, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s54, 0xff
; SI-NEXT:    s_lshl_b32 s8, s74, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x48, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s59, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v34
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v62, 60
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s46, 0xff
; SI-NEXT:    v_readlane_b32 s9, v62, 61
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v62, 62
; SI-NEXT:    v_readlane_b32 s9, v62, 63
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s26, v61, 0
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s26, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s68, 0xff
; SI-NEXT:    s_lshl_b32 s8, s22, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x50, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s45, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v31
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v61, 2
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s40, 0xff
; SI-NEXT:    v_readlane_b32 s9, v61, 3
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    v_readlane_b32 s27, v61, 1
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v61, 4
; SI-NEXT:    v_readlane_b32 s9, v61, 5
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s26, v61, 6
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s26, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s56, 0xff
; SI-NEXT:    s_lshl_b32 s8, s21, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x58, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s29, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v30
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v61, 8
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s24, 0xff
; SI-NEXT:    v_readlane_b32 s9, v61, 9
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v61, 10
; SI-NEXT:    v_readlane_b32 s9, v61, 11
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s20, v61, 12
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s20, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s34, 0xff
; SI-NEXT:    s_lshl_b32 s8, s18, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x60, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s23, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v29
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v61, 14
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s16, 0xff
; SI-NEXT:    v_readlane_b32 s9, v61, 15
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v61, 16
; SI-NEXT:    v_readlane_b32 s9, v61, 17
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s16, v61, 18
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s16, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s36, 0xff
; SI-NEXT:    s_lshl_b32 s8, s15, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x68, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    s_and_b32 s8, s19, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v27
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v61, 20
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_and_b32 s5, s10, 0xff
; SI-NEXT:    v_readlane_b32 s9, v61, 21
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v61, 22
; SI-NEXT:    v_readlane_b32 s9, v61, 23
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v61, 24
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s10, 24
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    v_mov_b32_e32 v2, s5
; SI-NEXT:    s_and_b32 s5, s14, 0xff
; SI-NEXT:    s_lshl_b32 s8, s12, 8
; SI-NEXT:    s_or_b32 s5, s5, s8
; SI-NEXT:    v_readlane_b32 s8, v61, 26
; SI-NEXT:    v_readlane_b32 s9, v61, 27
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x70, v0
; SI-NEXT:    s_and_b32 s8, s9, 0xff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v26
; SI-NEXT:    v_or_b32_e32 v1, s8, v1
; SI-NEXT:    v_readlane_b32 s8, v61, 28
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    v_readlane_b32 s9, v61, 29
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    s_lshl_b32 s5, s8, 8
; SI-NEXT:    v_readlane_b32 s8, v61, 30
; SI-NEXT:    v_readlane_b32 s9, v61, 31
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s8, 0xff
; SI-NEXT:    v_readlane_b32 s8, v61, 32
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s8, s5
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x78, v0
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    s_and_b32 s4, s88, 0xff
; SI-NEXT:    s_lshl_b32 s5, s6, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s13, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    v_readlane_b32 s27, v61, 7
; SI-NEXT:    v_readlane_b32 s21, v61, 13
; SI-NEXT:    v_readlane_b32 s17, v61, 19
; SI-NEXT:    v_readlane_b32 s11, v61, 25
; SI-NEXT:    v_readlane_b32 s9, v61, 33
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
; SI-NEXT:    v_or_b32_e32 v1, s5, v1
; SI-NEXT:    v_or_b32_e32 v1, s4, v1
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v128i8_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v63, s30, 0
; VI-NEXT:    v_writelane_b32 v63, s31, 1
; VI-NEXT:    v_writelane_b32 v63, s34, 2
; VI-NEXT:    v_writelane_b32 v63, s35, 3
; VI-NEXT:    v_writelane_b32 v63, s36, 4
; VI-NEXT:    v_writelane_b32 v63, s37, 5
; VI-NEXT:    v_writelane_b32 v63, s38, 6
; VI-NEXT:    v_writelane_b32 v63, s39, 7
; VI-NEXT:    v_writelane_b32 v63, s48, 8
; VI-NEXT:    v_writelane_b32 v63, s49, 9
; VI-NEXT:    v_writelane_b32 v63, s50, 10
; VI-NEXT:    v_writelane_b32 v63, s51, 11
; VI-NEXT:    v_writelane_b32 v63, s52, 12
; VI-NEXT:    v_writelane_b32 v63, s53, 13
; VI-NEXT:    v_writelane_b32 v63, s54, 14
; VI-NEXT:    v_writelane_b32 v63, s55, 15
; VI-NEXT:    v_writelane_b32 v63, s64, 16
; VI-NEXT:    v_writelane_b32 v63, s65, 17
; VI-NEXT:    v_writelane_b32 v63, s66, 18
; VI-NEXT:    v_writelane_b32 v63, s67, 19
; VI-NEXT:    v_writelane_b32 v63, s68, 20
; VI-NEXT:    v_writelane_b32 v63, s69, 21
; VI-NEXT:    v_writelane_b32 v63, s70, 22
; VI-NEXT:    v_writelane_b32 v63, s71, 23
; VI-NEXT:    v_writelane_b32 v63, s80, 24
; VI-NEXT:    v_writelane_b32 v63, s81, 25
; VI-NEXT:    v_writelane_b32 v63, s82, 26
; VI-NEXT:    v_writelane_b32 v63, s83, 27
; VI-NEXT:    v_writelane_b32 v63, s84, 28
; VI-NEXT:    v_writelane_b32 v63, s85, 29
; VI-NEXT:    v_writelane_b32 v63, s86, 30
; VI-NEXT:    v_readfirstlane_b32 s4, v19
; VI-NEXT:    v_writelane_b32 v63, s87, 31
; VI-NEXT:    v_readfirstlane_b32 s7, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s9, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s11, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s13, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s15, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s41, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s43, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s45, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s5, v2
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s4, v1
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; VI-NEXT:    s_cbranch_scc0 .LBB91_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 0
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 1
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s29, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s28, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s27, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s26, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s25, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s24, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s23, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s23, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s23, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s22, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s21, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s21, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s20, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s20, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s19, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s19, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s19, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s18, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s18, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s17, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s17, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s17, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s16, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s16, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s41, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s41, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s40, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s40, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s43, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s43, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s42, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s42, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 57
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    s_lshr_b32 s69, s7, 24
; VI-NEXT:    s_lshr_b32 s68, s7, 16
; VI-NEXT:    s_lshr_b32 s67, s7, 8
; VI-NEXT:    s_lshr_b32 s71, s6, 16
; VI-NEXT:    s_lshr_b32 s70, s6, 8
; VI-NEXT:    s_lshr_b32 s64, s9, 24
; VI-NEXT:    s_lshr_b32 s55, s9, 16
; VI-NEXT:    s_lshr_b32 s54, s9, 8
; VI-NEXT:    s_lshr_b32 s66, s8, 16
; VI-NEXT:    s_lshr_b32 s65, s8, 8
; VI-NEXT:    s_lshr_b32 s53, s11, 24
; VI-NEXT:    s_lshr_b32 s52, s11, 16
; VI-NEXT:    s_lshr_b32 s51, s11, 8
; VI-NEXT:    s_lshr_b32 s87, s10, 16
; VI-NEXT:    s_lshr_b32 s86, s10, 8
; VI-NEXT:    s_lshr_b32 s50, s13, 24
; VI-NEXT:    s_lshr_b32 s83, s13, 16
; VI-NEXT:    s_lshr_b32 s82, s13, 8
; VI-NEXT:    s_lshr_b32 s85, s12, 16
; VI-NEXT:    s_lshr_b32 s84, s12, 8
; VI-NEXT:    s_lshr_b32 s81, s14, 16
; VI-NEXT:    s_lshr_b32 s80, s14, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 54
; VI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; VI-NEXT:    s_lshr_b64 s[56:57], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[16:17], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[44:45], 24
; VI-NEXT:    s_cbranch_execnz .LBB91_4
; VI-NEXT:  .LBB91_2: ; %cmp.true
; VI-NEXT:    s_lshl_b32 s46, s45, 16
; VI-NEXT:    v_mov_b32_e32 v2, 0x40c00000
; VI-NEXT:    v_add_f32_e32 v1, s46, v2
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s45, s45, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s45, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[3:4]
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    s_lshl_b32 s45, s44, 16
; VI-NEXT:    v_add_f32_e32 v1, s45, v2
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s44, s44, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s44, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_lshrrev_b64 v[3:4], 16, v[3:4]
; VI-NEXT:    s_lshl_b32 s44, s43, 16
; VI-NEXT:    v_add_f32_e32 v1, s44, v2
; VI-NEXT:    s_and_b32 s43, s43, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v4, v6
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s43, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[3:4]
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    s_lshl_b32 s43, s42, 16
; VI-NEXT:    v_add_f32_e32 v1, s43, v2
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s42, s42, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s42, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s42, s41, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s42, v2
; VI-NEXT:    v_lshrrev_b64 v[59:60], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s41, s41, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s41, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s41, s40, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s41, v2
; VI-NEXT:    s_and_b32 s40, s40, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v60, v6
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s40, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s40, s15, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s40, v2
; VI-NEXT:    v_lshrrev_b64 v[14:15], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s15, s15, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s15, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s15, s14, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s15, v2
; VI-NEXT:    v_lshrrev_b64 v[16:17], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s14, s14, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s14, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s14, s13, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s14, v2
; VI-NEXT:    v_lshrrev_b64 v[17:18], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s13, s13, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s13, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s13, s12, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s13, v2
; VI-NEXT:    v_lshrrev_b64 v[55:56], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s12, s12, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s12, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s12, s11, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s12, v2
; VI-NEXT:    v_lshrrev_b64 v[21:22], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s11, s11, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s11, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s11, s10, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s11, v2
; VI-NEXT:    v_lshrrev_b64 v[57:58], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s10, s10, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s10, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s10, s9, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s10, v2
; VI-NEXT:    v_lshrrev_b64 v[40:41], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s9, s9, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s9, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s9, s8, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s9, v2
; VI-NEXT:    v_lshrrev_b64 v[12:13], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s8, s8, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s8, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s8, s7, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s8, v2
; VI-NEXT:    v_lshrrev_b64 v[42:43], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s7, s7, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s7, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s7, s6, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s7, v2
; VI-NEXT:    v_lshrrev_b64 v[8:9], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s6, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s17, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[9:10], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s17, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s16, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[46:47], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s16, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s19, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[26:27], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s19, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s18, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[24:25], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s18, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s21, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[29:30], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s21, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s20, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[31:32], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s20, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s23, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[32:33], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s23, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s22, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[34:35], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s22, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s25, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[35:36], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s25, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s24, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[37:38], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s24, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s27, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[38:39], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s27, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s26, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[48:49], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s26, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s29, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[49:50], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s29, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s28, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[51:52], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s6, s5, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s6, v2
; VI-NEXT:    v_lshrrev_b64 v[52:53], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s5, s5, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v1, s5, v2
; VI-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v4, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v4, v5, vcc
; VI-NEXT:    s_lshl_b32 s5, s4, 16
; VI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; VI-NEXT:    v_add_f32_e32 v1, s5, v2
; VI-NEXT:    v_lshrrev_b64 v[19:20], 16, v[3:4]
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_and_b32 s4, s4, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_add_f32_e32 v2, s4, v2
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
; VI-NEXT:    v_bfe_u32 v3, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_lshrrev_b64 v[44:45], 16, v[1:2]
; VI-NEXT:    v_mov_b32_e32 v45, v19
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[44:45]
; VI-NEXT:    v_mov_b32_e32 v53, v51
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[52:53]
; VI-NEXT:    v_mov_b32_e32 v50, v48
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v39, v37
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[49:50]
; VI-NEXT:    v_lshrrev_b64 v[2:3], 24, v[38:39]
; VI-NEXT:    v_mov_b32_e32 v36, v34
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v33, v31
; VI-NEXT:    v_lshrrev_b64 v[3:4], 24, v[35:36]
; VI-NEXT:    v_mov_b32_e32 v30, v24
; VI-NEXT:    v_lshrrev_b64 v[4:5], 24, v[32:33]
; VI-NEXT:    v_mov_b32_e32 v15, v6
; VI-NEXT:    v_mov_b32_e32 v27, v46
; VI-NEXT:    v_lshrrev_b64 v[5:6], 24, v[29:30]
; VI-NEXT:    v_lshrrev_b64 v[6:7], 24, v[26:27]
; VI-NEXT:    v_mov_b32_e32 v7, v8
; VI-NEXT:    v_mov_b32_e32 v10, v8
; VI-NEXT:    v_lshrrev_b32_e32 v30, 24, v46
; VI-NEXT:    v_lshrrev_b32_e32 v25, 16, v46
; VI-NEXT:    v_mov_b32_e32 v47, v46
; VI-NEXT:    v_lshrrev_b32_e32 v50, 8, v46
; VI-NEXT:    v_lshrrev_b32_e32 v45, 24, v8
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v8
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[9:10]
; VI-NEXT:    v_mov_b32_e32 v43, v12
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v9
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[42:43]
; VI-NEXT:    v_mov_b32_e32 v41, v57
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[40:41]
; VI-NEXT:    v_mov_b32_e32 v22, v55
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[21:22]
; VI-NEXT:    v_mov_b32_e32 v18, v16
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[17:18]
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[14:15]
; VI-NEXT:    v_mov_b32_e32 v18, v14
; VI-NEXT:    v_lshrrev_b32_e32 v2, 24, v19
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v19
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v44
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v44
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v51
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v51
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v52
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v52
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 24, v48
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v48
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v48
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v49
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v49
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 24, v37
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v37
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v37
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v38
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v38
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 24, v34
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v34
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[59:60]
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v34
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v35
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v35
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 24, v31
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v31
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v32
; VI-NEXT:    v_lshrrev_b32_e32 v22, 24, v51
; VI-NEXT:    v_mov_b32_e32 v27, v34
; VI-NEXT:    v_mov_b32_e32 v60, v31
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v31
; VI-NEXT:    v_lshrrev_b32_e32 v43, 8, v8
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v9
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 24, v12
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v12
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v12
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v12
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v42
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v42
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v42
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 24, v57
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v57
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v57
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v57
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v40
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v40
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v40
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 24, v55
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v55
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v55
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v55
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v21
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v21
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 24, v16
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v16
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v16
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v17
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v17
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_mov_b32_e32 v7, v12
; VI-NEXT:    v_lshrrev_b32_e32 v12, 24, v7
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v12, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, v16
; VI-NEXT:    v_mov_b32_e32 v16, v17
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v12, v7
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_lshrrev_b32_e32 v13, 8, v7
; VI-NEXT:    v_lshrrev_b32_e32 v7, 16, v14
; VI-NEXT:    v_lshrrev_b32_e32 v17, 8, v14
; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    v_lshrrev_b64 v[10:11], 24, v[53:54]
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v34, 8, v32
; VI-NEXT:    v_lshrrev_b32_e32 v28, 24, v24
; VI-NEXT:    v_lshrrev_b32_e32 v11, 16, v24
; VI-NEXT:    v_mov_b32_e32 v23, v24
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v24
; VI-NEXT:    v_lshrrev_b32_e32 v24, 16, v29
; VI-NEXT:    v_lshrrev_b32_e32 v2, 8, v29
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v26
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v26
; VI-NEXT:    v_lshrrev_b32_e32 v61, 16, v9
; VI-NEXT:    v_mov_b32_e32 v42, v21
; VI-NEXT:    v_lshrrev_b32_e32 v21, 16, v59
; VI-NEXT:    v_mov_b32_e32 v41, v59
; VI-NEXT:    v_lshrrev_b32_e32 v9, 8, v59
; VI-NEXT:    v_mov_b32_e32 v59, v53
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_mov_b32_e32 v15, v14
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_mov_b32_e32 v36, v55
; VI-NEXT:    v_lshrrev_b32_e32 v54, 24, v15
; VI-NEXT:    v_lshrrev_b32_e32 v39, 16, v15
; VI-NEXT:    v_lshrrev_b32_e32 v40, 8, v15
; VI-NEXT:    v_lshrrev_b32_e32 v15, 24, v36
; VI-NEXT:    v_lshrrev_b32_e32 v55, 16, v36
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_mov_b32_e32 v56, v36
; VI-NEXT:    v_lshrrev_b32_e32 v57, 8, v36
; VI-NEXT:    v_lshrrev_b32_e32 v36, 16, v53
; VI-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
; VI-NEXT:    s_branch .LBB91_5
; VI-NEXT:  .LBB91_3:
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr80
; VI-NEXT:    ; implicit-def: $sgpr81
; VI-NEXT:    ; implicit-def: $sgpr84
; VI-NEXT:    ; implicit-def: $sgpr85
; VI-NEXT:    ; implicit-def: $sgpr82
; VI-NEXT:    ; implicit-def: $sgpr83
; VI-NEXT:    ; implicit-def: $sgpr50
; VI-NEXT:    ; implicit-def: $sgpr86
; VI-NEXT:    ; implicit-def: $sgpr87
; VI-NEXT:    ; implicit-def: $sgpr51
; VI-NEXT:    ; implicit-def: $sgpr52
; VI-NEXT:    ; implicit-def: $sgpr53
; VI-NEXT:    ; implicit-def: $sgpr65
; VI-NEXT:    ; implicit-def: $sgpr66
; VI-NEXT:    ; implicit-def: $sgpr54
; VI-NEXT:    ; implicit-def: $sgpr55
; VI-NEXT:    ; implicit-def: $sgpr64
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr71
; VI-NEXT:    ; implicit-def: $sgpr67
; VI-NEXT:    ; implicit-def: $sgpr68
; VI-NEXT:    ; implicit-def: $sgpr69
; VI-NEXT:    ; implicit-def: $sgpr76
; VI-NEXT:    ; implicit-def: $sgpr74
; VI-NEXT:    ; implicit-def: $sgpr72
; VI-NEXT:    ; implicit-def: $sgpr62
; VI-NEXT:    ; implicit-def: $sgpr60
; VI-NEXT:    ; implicit-def: $sgpr58
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr48
; VI-NEXT:    ; implicit-def: $sgpr38
; VI-NEXT:    ; implicit-def: $sgpr36
; VI-NEXT:    ; implicit-def: $sgpr34
; VI-NEXT:    ; implicit-def: $sgpr30
; VI-NEXT:    ; implicit-def: $sgpr90
; VI-NEXT:    ; implicit-def: $sgpr88
; VI-NEXT:    ; implicit-def: $sgpr78
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    s_branch .LBB91_2
; VI-NEXT:  .LBB91_4:
; VI-NEXT:    v_readlane_b32 s47, v62, 0
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 1
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 2
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 3
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 4
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 5
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 6
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 7
; VI-NEXT:    v_mov_b32_e32 v7, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 9
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 10
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 11
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 12
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 13
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 14
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 15
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 17
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 18
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 19
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 20
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 21
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 22
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 23
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 24
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 25
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 26
; VI-NEXT:    v_mov_b32_e32 v34, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 27
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 28
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s70
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s66
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s65
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s64
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s55
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s54
; VI-NEXT:    v_readlane_b32 s47, v62, 29
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s87
; VI-NEXT:    v_mov_b32_e32 v10, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 30
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s86
; VI-NEXT:    v_mov_b32_e32 v24, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 31
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s53
; VI-NEXT:    v_mov_b32_e32 v2, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 32
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s52
; VI-NEXT:    v_mov_b32_e32 v28, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 33
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s51
; VI-NEXT:    v_mov_b32_e32 v11, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 34
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s85
; VI-NEXT:    v_mov_b32_e32 v31, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 35
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s84
; VI-NEXT:    v_mov_b32_e32 v33, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 36
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s50
; VI-NEXT:    v_mov_b32_e32 v58, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 37
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s83
; VI-NEXT:    v_mov_b32_e32 v30, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 38
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s82
; VI-NEXT:    v_mov_b32_e32 v25, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 39
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s81
; VI-NEXT:    v_mov_b32_e32 v50, s47
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s80
; VI-NEXT:    v_readlane_b32 s47, v62, 40
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 41
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 42
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 43
; VI-NEXT:    v_mov_b32_e32 v20, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 44
; VI-NEXT:    v_mov_b32_e32 v17, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 45
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 46
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s7
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s9
; VI-NEXT:    v_mov_b32_e32 v8, s15
; VI-NEXT:    v_mov_b32_e32 v22, v7
; VI-NEXT:    v_mov_b32_e32 v7, s90
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s10
; VI-NEXT:    v_mov_b32_e32 v7, s30
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s11
; VI-NEXT:    v_mov_b32_e32 v7, s34
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s13
; VI-NEXT:    v_mov_b32_e32 v7, s36
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s46
; VI-NEXT:    v_mov_b32_e32 v7, s38
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_readlane_b32 s47, v62, 47
; VI-NEXT:    v_mov_b32_e32 v13, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 48
; VI-NEXT:    v_mov_b32_e32 v21, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 49
; VI-NEXT:    v_mov_b32_e32 v9, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 50
; VI-NEXT:    v_mov_b32_e32 v54, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 51
; VI-NEXT:    v_mov_b32_e32 v39, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 52
; VI-NEXT:    v_mov_b32_e32 v40, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 53
; VI-NEXT:    v_mov_b32_e32 v36, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 54
; VI-NEXT:    v_mov_b32_e32 v53, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 55
; VI-NEXT:    v_mov_b32_e32 v46, s68
; VI-NEXT:    v_mov_b32_e32 v15, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 56
; VI-NEXT:    v_mov_b32_e32 v44, s4
; VI-NEXT:    v_mov_b32_e32 v1, s56
; VI-NEXT:    v_mov_b32_e32 v3, s60
; VI-NEXT:    v_mov_b32_e32 v45, s78
; VI-NEXT:    v_mov_b32_e32 v43, s88
; VI-NEXT:    v_mov_b32_e32 v7, s48
; VI-NEXT:    v_mov_b32_e32 v55, s47
; VI-NEXT:    v_readlane_b32 s47, v62, 57
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v61, s71
; VI-NEXT:    v_mov_b32_e32 v57, s47
; VI-NEXT:    v_mov_b32_e32 v19, s5
; VI-NEXT:    v_mov_b32_e32 v52, s28
; VI-NEXT:    v_mov_b32_e32 v51, s29
; VI-NEXT:    v_mov_b32_e32 v49, s26
; VI-NEXT:    v_mov_b32_e32 v48, s27
; VI-NEXT:    v_mov_b32_e32 v38, s24
; VI-NEXT:    v_mov_b32_e32 v37, s25
; VI-NEXT:    v_mov_b32_e32 v35, s22
; VI-NEXT:    v_mov_b32_e32 v27, s23
; VI-NEXT:    v_mov_b32_e32 v32, s20
; VI-NEXT:    v_mov_b32_e32 v60, s21
; VI-NEXT:    v_mov_b32_e32 v29, s18
; VI-NEXT:    v_mov_b32_e32 v23, s19
; VI-NEXT:    v_mov_b32_e32 v26, s16
; VI-NEXT:    v_mov_b32_e32 v47, s17
; VI-NEXT:    v_mov_b32_e32 v42, s12
; VI-NEXT:    v_mov_b32_e32 v16, s14
; VI-NEXT:    v_mov_b32_e32 v18, s40
; VI-NEXT:    v_mov_b32_e32 v12, s41
; VI-NEXT:    v_mov_b32_e32 v41, s42
; VI-NEXT:    v_mov_b32_e32 v14, s43
; VI-NEXT:    v_mov_b32_e32 v59, s44
; VI-NEXT:    v_mov_b32_e32 v56, s45
; VI-NEXT:    v_mov_b32_e32 v1, s58
; VI-NEXT:    v_mov_b32_e32 v3, s62
; VI-NEXT:    v_mov_b32_e32 v4, s72
; VI-NEXT:    v_mov_b32_e32 v5, s74
; VI-NEXT:    v_mov_b32_e32 v6, s76
; VI-NEXT:    v_mov_b32_e32 v45, s69
; VI-NEXT:    v_mov_b32_e32 v43, s67
; VI-NEXT:    v_mov_b32_e32 v7, v20
; VI-NEXT:    v_mov_b32_e32 v20, v10
; VI-NEXT:  .LBB91_5: ; %end
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v6, v33, v6, s4
; VI-NEXT:    v_perm_b32 v26, v26, v58, s4
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v6, v26, v6
; VI-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v6, v25, v30, s4
; VI-NEXT:    v_perm_b32 v33, v47, v50, s4
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v6, v33, v6
; VI-NEXT:    v_add_u32_e32 v25, vcc, 4, v0
; VI-NEXT:    v_perm_b32 v5, v24, v5, s4
; VI-NEXT:    buffer_store_dword v6, v25, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v6, v29, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v5, v6, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v6, v11, v28, s4
; VI-NEXT:    v_perm_b32 v5, v23, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_add_u32_e32 v6, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v5, v32, v34, s4
; VI-NEXT:    v_readlane_b32 s87, v63, 31
; VI-NEXT:    v_readlane_b32 s86, v63, 30
; VI-NEXT:    v_readlane_b32 s85, v63, 29
; VI-NEXT:    v_readlane_b32 s84, v63, 28
; VI-NEXT:    v_readlane_b32 s83, v63, 27
; VI-NEXT:    v_readlane_b32 s82, v63, 26
; VI-NEXT:    v_readlane_b32 s81, v63, 25
; VI-NEXT:    v_readlane_b32 s80, v63, 24
; VI-NEXT:    v_readlane_b32 s71, v63, 23
; VI-NEXT:    v_readlane_b32 s70, v63, 22
; VI-NEXT:    v_readlane_b32 s69, v63, 21
; VI-NEXT:    v_readlane_b32 s68, v63, 20
; VI-NEXT:    v_readlane_b32 s67, v63, 19
; VI-NEXT:    v_readlane_b32 s66, v63, 18
; VI-NEXT:    v_readlane_b32 s65, v63, 17
; VI-NEXT:    v_readlane_b32 s64, v63, 16
; VI-NEXT:    v_readlane_b32 s55, v63, 15
; VI-NEXT:    v_readlane_b32 s54, v63, 14
; VI-NEXT:    v_readlane_b32 s53, v63, 13
; VI-NEXT:    v_readlane_b32 s52, v63, 12
; VI-NEXT:    v_readlane_b32 s51, v63, 11
; VI-NEXT:    v_readlane_b32 s50, v63, 10
; VI-NEXT:    v_readlane_b32 s49, v63, 9
; VI-NEXT:    v_readlane_b32 s48, v63, 8
; VI-NEXT:    v_readlane_b32 s39, v63, 7
; VI-NEXT:    v_readlane_b32 s38, v63, 6
; VI-NEXT:    v_readlane_b32 s37, v63, 5
; VI-NEXT:    v_readlane_b32 s36, v63, 4
; VI-NEXT:    v_readlane_b32 s35, v63, 3
; VI-NEXT:    v_readlane_b32 s34, v63, 2
; VI-NEXT:    v_readlane_b32 s31, v63, 1
; VI-NEXT:    v_readlane_b32 s30, v63, 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v2, v4, s4
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v4, v5, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v4, v60, v20, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v5, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_add_u32_e32 v5, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v35, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v3, v4, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v27, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v4, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_add_u32_e32 v4, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v38, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v4, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v2, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v37, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v49, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v3, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v48, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v52, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v51, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v22, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v44, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v19, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v59, v53, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v36, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v55, v15, s4
; VI-NEXT:    v_perm_b32 v1, v56, v57, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v41, v9, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v21, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, v39, v54, s4
; VI-NEXT:    v_perm_b32 v1, v14, v40, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v18, v17, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v7, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v12, v13, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v16, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v8, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v42, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v61, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v46, v45, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    v_perm_b32 v1, v1, v43, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v128i8_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v63, s30, 0
; GFX9-NEXT:    v_writelane_b32 v63, s31, 1
; GFX9-NEXT:    v_writelane_b32 v63, s34, 2
; GFX9-NEXT:    v_writelane_b32 v63, s35, 3
; GFX9-NEXT:    v_writelane_b32 v63, s36, 4
; GFX9-NEXT:    v_writelane_b32 v63, s37, 5
; GFX9-NEXT:    v_writelane_b32 v63, s38, 6
; GFX9-NEXT:    v_writelane_b32 v63, s39, 7
; GFX9-NEXT:    v_writelane_b32 v63, s48, 8
; GFX9-NEXT:    v_writelane_b32 v63, s49, 9
; GFX9-NEXT:    v_writelane_b32 v63, s50, 10
; GFX9-NEXT:    v_writelane_b32 v63, s51, 11
; GFX9-NEXT:    v_writelane_b32 v63, s52, 12
; GFX9-NEXT:    v_writelane_b32 v63, s53, 13
; GFX9-NEXT:    v_writelane_b32 v63, s54, 14
; GFX9-NEXT:    v_writelane_b32 v63, s55, 15
; GFX9-NEXT:    v_writelane_b32 v63, s64, 16
; GFX9-NEXT:    v_writelane_b32 v63, s65, 17
; GFX9-NEXT:    v_writelane_b32 v63, s66, 18
; GFX9-NEXT:    v_writelane_b32 v63, s67, 19
; GFX9-NEXT:    v_writelane_b32 v63, s68, 20
; GFX9-NEXT:    v_writelane_b32 v63, s69, 21
; GFX9-NEXT:    v_writelane_b32 v63, s70, 22
; GFX9-NEXT:    v_writelane_b32 v63, s71, 23
; GFX9-NEXT:    v_writelane_b32 v63, s80, 24
; GFX9-NEXT:    v_writelane_b32 v63, s81, 25
; GFX9-NEXT:    v_writelane_b32 v63, s82, 26
; GFX9-NEXT:    v_writelane_b32 v63, s83, 27
; GFX9-NEXT:    v_writelane_b32 v63, s84, 28
; GFX9-NEXT:    v_writelane_b32 v63, s85, 29
; GFX9-NEXT:    v_writelane_b32 v63, s86, 30
; GFX9-NEXT:    v_writelane_b32 v63, s87, 31
; GFX9-NEXT:    v_writelane_b32 v63, s96, 32
; GFX9-NEXT:    v_writelane_b32 v63, s97, 33
; GFX9-NEXT:    v_writelane_b32 v63, s98, 34
; GFX9-NEXT:    v_readfirstlane_b32 s4, v19
; GFX9-NEXT:    v_writelane_b32 v63, s99, 35
; GFX9-NEXT:    v_readfirstlane_b32 s7, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s9, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s11, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s13, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s15, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s41, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s43, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s45, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s5, v2
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s4, v1
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; GFX9-NEXT:    s_cbranch_scc0 .LBB91_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 0
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s29, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s28, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s27, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s27, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s26, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s26, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s25, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s25, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s25, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s24, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s24, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s23, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s23, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s23, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s22, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s22, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s21, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s21, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s21, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s20, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s20, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s19, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s19, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s19, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s18, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s18, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s17, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s17, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s17, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s16, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s16, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s41, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s41, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s43, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s43, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s42, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s42, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s45, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s45, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 49
; GFX9-NEXT:    s_lshr_b32 s46, s44, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s44, 8
; GFX9-NEXT:    s_lshr_b32 s52, s7, 24
; GFX9-NEXT:    s_lshr_b32 s67, s7, 16
; GFX9-NEXT:    s_lshr_b32 s55, s7, 8
; GFX9-NEXT:    s_lshr_b32 s65, s6, 16
; GFX9-NEXT:    s_lshr_b32 s64, s6, 8
; GFX9-NEXT:    s_lshr_b32 s48, s9, 24
; GFX9-NEXT:    s_lshr_b32 s68, s9, 16
; GFX9-NEXT:    s_lshr_b32 s51, s9, 8
; GFX9-NEXT:    s_lshr_b32 s54, s8, 16
; GFX9-NEXT:    s_lshr_b32 s53, s8, 8
; GFX9-NEXT:    s_lshr_b32 s98, s11, 24
; GFX9-NEXT:    s_lshr_b32 s69, s11, 16
; GFX9-NEXT:    s_lshr_b32 s39, s11, 8
; GFX9-NEXT:    s_lshr_b32 s50, s10, 16
; GFX9-NEXT:    s_lshr_b32 s49, s10, 8
; GFX9-NEXT:    s_lshr_b32 s86, s13, 24
; GFX9-NEXT:    s_lshr_b32 s70, s13, 16
; GFX9-NEXT:    s_lshr_b32 s97, s13, 8
; GFX9-NEXT:    s_lshr_b32 s38, s12, 16
; GFX9-NEXT:    s_lshr_b32 s99, s12, 8
; GFX9-NEXT:    s_lshr_b32 s82, s15, 24
; GFX9-NEXT:    s_lshr_b32 s71, s15, 16
; GFX9-NEXT:    s_lshr_b32 s85, s15, 8
; GFX9-NEXT:    s_lshr_b32 s96, s14, 16
; GFX9-NEXT:    s_lshr_b32 s87, s14, 8
; GFX9-NEXT:    s_lshr_b32 s80, s41, 16
; GFX9-NEXT:    s_lshr_b32 s84, s40, 16
; GFX9-NEXT:    s_lshr_b32 s83, s40, 8
; GFX9-NEXT:    s_lshr_b32 s81, s43, 16
; GFX9-NEXT:    s_lshr_b32 s66, s45, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 47
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[16:17], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[44:45], 24
; GFX9-NEXT:    s_cbranch_execnz .LBB91_4
; GFX9-NEXT:  .LBB91_2: ; %cmp.true
; GFX9-NEXT:    s_and_b32 s46, s45, 0xffff0000
; GFX9-NEXT:    v_mov_b32_e32 v8, 0x40c00000
; GFX9-NEXT:    v_add_f32_e32 v1, s46, v8
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_lshl_b32 s45, s45, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v1, s45, v8
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX9-NEXT:    s_and_b32 s45, s44, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v2, v4, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v1, s45, v8
; GFX9-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v3, v3, v1
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v3, 0x7fff, v3
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    s_lshl_b32 s44, s44, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
; GFX9-NEXT:    v_add_f32_e32 v3, s44, v8
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    s_and_b32 s44, s43, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v3, s44, v8
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    s_lshl_b32 s43, s43, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v3, s43, v8
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    s_and_b32 s43, s42, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v4, v6, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v3, s43, v8
; GFX9-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v5, v5, v3
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v5, 0x7fff, v5
; GFX9-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    s_lshl_b32 s42, s42, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc
; GFX9-NEXT:    v_add_f32_e32 v5, s42, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s42, s41, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v3, v3, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s42, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s41, s41, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s41, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s41, s40, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v12, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s41, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s40, s40, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s40, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s40, s15, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v11, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s40, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s15, s15, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s15, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s15, s14, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v14, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s15, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s14, s14, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s14, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s14, s13, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v13, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s14, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s13, s13, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s13, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s13, s12, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v16, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s13, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s12, s12, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s12, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s12, s11, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v15, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s12, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s11, s11, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s11, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s11, s10, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v20, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s11, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s10, s10, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s10, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s10, s9, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v19, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s10, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s9, s9, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s9, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s9, s8, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v22, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s9, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s8, s8, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s8, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s8, s7, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v21, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s8, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s7, s7, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s7, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s7, s6, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v24, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s7, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s6, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s17, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v23, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s6, s17, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s6, s16, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v31, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s16, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s19, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v30, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s6, s19, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s6, s18, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v49, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s18, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s21, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v48, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s6, s21, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s6, s20, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v53, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s20, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s23, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v52, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s6, s23, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s6, s22, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v55, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s22, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s25, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v54, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s6, s25, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s6, s24, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v41, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s24, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s27, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v40, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s6, s27, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s6, s26, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v43, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s26, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s29, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v42, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s6, s29, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s6, s28, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v45, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s6, s28, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s6, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v9, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    s_and_b32 s6, s5, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v44, v5, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v5, s6, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    s_lshl_b32 s5, s5, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s5, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    s_and_b32 s5, s4, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v47, v9, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v5, s5, v8
; GFX9-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v6, v6, v5
; GFX9-NEXT:    v_add_u32_e32 v6, 0x7fff, v6
; GFX9-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    s_lshl_b32 s4, s4, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v6, v7, vcc
; GFX9-NEXT:    v_add_f32_e32 v6, s4, v8
; GFX9-NEXT:    v_bfe_u32 v7, v6, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v7, v7, v6
; GFX9-NEXT:    v_add_u32_e32 v7, 0x7fff, v7
; GFX9-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v7, v8, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v8, 16, v6
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v8
; GFX9-NEXT:    v_lshl_or_b32 v46, v5, 16, v6
; GFX9-NEXT:    v_lshrrev_b64 v[5:6], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[5:6], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[5:6], 24, v[19:20]
; GFX9-NEXT:    v_lshrrev_b64 v[34:35], 24, v[42:43]
; GFX9-NEXT:    v_lshrrev_b64 v[35:36], 24, v[40:41]
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[36:37], 24, v[54:55]
; GFX9-NEXT:    v_lshrrev_b64 v[5:6], 24, v[15:16]
; GFX9-NEXT:    v_lshrrev_b64 v[37:38], 24, v[52:53]
; GFX9-NEXT:    v_lshrrev_b32_e32 v7, 16, v52
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v52
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 24, v24
; GFX9-NEXT:    v_lshrrev_b32_e32 v24, 8, v24
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; GFX9-NEXT:    v_lshrrev_b32_e32 v23, 8, v23
; GFX9-NEXT:    v_lshrrev_b64 v[5:6], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v23, 24, v22
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v22
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v21, 8, v21
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v21, 24, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 8, v20
; GFX9-NEXT:    v_lshrrev_b64 v[5:6], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 8, v19
; GFX9-NEXT:    v_lshrrev_b64 v[25:26], 24, v[46:47]
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 24, v16
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
; GFX9-NEXT:    v_lshrrev_b64 v[28:29], 24, v[44:45]
; GFX9-NEXT:    v_lshrrev_b64 v[38:39], 24, v[48:49]
; GFX9-NEXT:    v_lshrrev_b64 v[50:51], 24, v[30:31]
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[26:27], 24, v[3:4]
; GFX9-NEXT:    v_lshrrev_b64 v[32:33], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 24, v47
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v47
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v46
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v46
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 24, v45
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v45
; GFX9-NEXT:    v_lshrrev_b32_e32 v29, 16, v44
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v44
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 24, v43
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 8, v43
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 16, v42
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 8, v42
; GFX9-NEXT:    v_lshrrev_b32_e32 v10, 24, v41
; GFX9-NEXT:    v_lshrrev_b32_e32 v41, 8, v41
; GFX9-NEXT:    v_lshrrev_b32_e32 v6, 16, v40
; GFX9-NEXT:    v_lshrrev_b32_e32 v40, 8, v40
; GFX9-NEXT:    v_lshrrev_b32_e32 v27, 24, v55
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 8, v55
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v54
; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 8, v54
; GFX9-NEXT:    v_lshrrev_b32_e32 v5, 24, v53
; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 8, v53
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 8, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v49, 16, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 8, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v48, 24, v31
; GFX9-NEXT:    v_lshrrev_b32_e32 v31, 8, v31
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 16, v30
; GFX9-NEXT:    v_lshrrev_b32_e32 v30, 8, v30
; GFX9-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
; GFX9-NEXT:    v_lshrrev_b32_e32 v24, 24, v14
; GFX9-NEXT:    v_lshrrev_b32_e32 v14, 8, v14
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 16, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v13, 8, v13
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 24, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v12, 8, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 16, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
; GFX9-NEXT:    v_lshrrev_b32_e32 v21, 24, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v4, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v23, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v3, 8, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
; GFX9-NEXT:    s_branch .LBB91_5
; GFX9-NEXT:  .LBB91_3:
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr66
; GFX9-NEXT:    ; implicit-def: $sgpr81
; GFX9-NEXT:    ; implicit-def: $sgpr83
; GFX9-NEXT:    ; implicit-def: $sgpr84
; GFX9-NEXT:    ; implicit-def: $sgpr80
; GFX9-NEXT:    ; implicit-def: $sgpr87
; GFX9-NEXT:    ; implicit-def: $sgpr96
; GFX9-NEXT:    ; implicit-def: $sgpr85
; GFX9-NEXT:    ; implicit-def: $sgpr71
; GFX9-NEXT:    ; implicit-def: $sgpr82
; GFX9-NEXT:    ; implicit-def: $sgpr99
; GFX9-NEXT:    ; implicit-def: $sgpr38
; GFX9-NEXT:    ; implicit-def: $sgpr97
; GFX9-NEXT:    ; implicit-def: $sgpr70
; GFX9-NEXT:    ; implicit-def: $sgpr86
; GFX9-NEXT:    ; implicit-def: $sgpr49
; GFX9-NEXT:    ; implicit-def: $sgpr50
; GFX9-NEXT:    ; implicit-def: $sgpr39
; GFX9-NEXT:    ; implicit-def: $sgpr69
; GFX9-NEXT:    ; implicit-def: $sgpr98
; GFX9-NEXT:    ; implicit-def: $sgpr53
; GFX9-NEXT:    ; implicit-def: $sgpr54
; GFX9-NEXT:    ; implicit-def: $sgpr51
; GFX9-NEXT:    ; implicit-def: $sgpr68
; GFX9-NEXT:    ; implicit-def: $sgpr48
; GFX9-NEXT:    ; implicit-def: $sgpr64
; GFX9-NEXT:    ; implicit-def: $sgpr65
; GFX9-NEXT:    ; implicit-def: $sgpr55
; GFX9-NEXT:    ; implicit-def: $sgpr67
; GFX9-NEXT:    ; implicit-def: $sgpr52
; GFX9-NEXT:    ; implicit-def: $sgpr76
; GFX9-NEXT:    ; implicit-def: $sgpr74
; GFX9-NEXT:    ; implicit-def: $sgpr72
; GFX9-NEXT:    ; implicit-def: $sgpr62
; GFX9-NEXT:    ; implicit-def: $sgpr60
; GFX9-NEXT:    ; implicit-def: $sgpr58
; GFX9-NEXT:    ; implicit-def: $sgpr56
; GFX9-NEXT:    ; implicit-def: $sgpr36
; GFX9-NEXT:    ; implicit-def: $sgpr34
; GFX9-NEXT:    ; implicit-def: $sgpr30
; GFX9-NEXT:    ; implicit-def: $sgpr94
; GFX9-NEXT:    ; implicit-def: $sgpr92
; GFX9-NEXT:    ; implicit-def: $sgpr90
; GFX9-NEXT:    ; implicit-def: $sgpr88
; GFX9-NEXT:    ; implicit-def: $sgpr78
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    s_branch .LBB91_2
; GFX9-NEXT:  .LBB91_4:
; GFX9-NEXT:    v_mov_b32_e32 v1, s5
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s28
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s29
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s26
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s27
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s24
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s25
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s22
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s23
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s20
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s21
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s18
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s16
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s7
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s9
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s10
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s11
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s12
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s13
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s14
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s15
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s40
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s41
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s42
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s43
; GFX9-NEXT:    v_mov_b32_e32 v8, s4
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s44
; GFX9-NEXT:    v_readlane_b32 s4, v62, 0
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s45
; GFX9-NEXT:    v_mov_b32_e32 v9, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 1
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 2
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 3
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 4
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 5
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 6
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 7
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 8
; GFX9-NEXT:    v_mov_b32_e32 v18, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 9
; GFX9-NEXT:    v_mov_b32_e32 v46, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 10
; GFX9-NEXT:    v_mov_b32_e32 v57, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 11
; GFX9-NEXT:    v_mov_b32_e32 v47, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 12
; GFX9-NEXT:    v_mov_b32_e32 v29, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 13
; GFX9-NEXT:    v_mov_b32_e32 v44, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 14
; GFX9-NEXT:    v_mov_b32_e32 v58, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 15
; GFX9-NEXT:    v_mov_b32_e32 v45, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 16
; GFX9-NEXT:    v_mov_b32_e32 v56, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 17
; GFX9-NEXT:    v_mov_b32_e32 v42, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 18
; GFX9-NEXT:    v_mov_b32_e32 v59, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 19
; GFX9-NEXT:    v_mov_b32_e32 v43, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 20
; GFX9-NEXT:    v_mov_b32_e32 v6, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 21
; GFX9-NEXT:    v_mov_b32_e32 v40, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 22
; GFX9-NEXT:    v_mov_b32_e32 v10, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 23
; GFX9-NEXT:    v_mov_b32_e32 v41, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 24
; GFX9-NEXT:    v_mov_b32_e32 v61, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 25
; GFX9-NEXT:    v_mov_b32_e32 v54, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 26
; GFX9-NEXT:    v_mov_b32_e32 v27, s4
; GFX9-NEXT:    v_mov_b32_e32 v26, s78
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v26, s88
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s67
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s68
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s69
; GFX9-NEXT:    v_readlane_b32 s4, v62, 27
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s70
; GFX9-NEXT:    v_mov_b32_e32 v55, s4
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v26, s90
; GFX9-NEXT:    v_readlane_b32 s4, v62, 28
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s71
; GFX9-NEXT:    v_mov_b32_e32 v7, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 29
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s80
; GFX9-NEXT:    v_mov_b32_e32 v60, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 30
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v26, s92
; GFX9-NEXT:    v_mov_b32_e32 v1, s81
; GFX9-NEXT:    v_mov_b32_e32 v5, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 31
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s66
; GFX9-NEXT:    v_mov_b32_e32 v53, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 32
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v49, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 33
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v26, s94
; GFX9-NEXT:    v_mov_b32_e32 v1, s65
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 34
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s64
; GFX9-NEXT:    v_mov_b32_e32 v33, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 35
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s52
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readlane_b32 s4, v62, 36
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s55
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 37
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s54
; GFX9-NEXT:    v_mov_b32_e32 v30, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 38
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s53
; GFX9-NEXT:    v_mov_b32_e32 v48, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 39
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s48
; GFX9-NEXT:    v_mov_b32_e32 v31, s4
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s51
; GFX9-NEXT:    v_readlane_b32 s4, v62, 40
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s50
; GFX9-NEXT:    v_mov_b32_e32 v16, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 41
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s49
; GFX9-NEXT:    v_mov_b32_e32 v12, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 42
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s98
; GFX9-NEXT:    v_mov_b32_e32 v23, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 43
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s39
; GFX9-NEXT:    v_mov_b32_e32 v3, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 44
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s38
; GFX9-NEXT:    v_mov_b32_e32 v21, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 45
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s86
; GFX9-NEXT:    v_mov_b32_e32 v4, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 46
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s97
; GFX9-NEXT:    v_mov_b32_e32 v22, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 47
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v1, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 48
; GFX9-NEXT:    v_mov_b32_e32 v26, s30
; GFX9-NEXT:    v_mov_b32_e32 v20, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 49
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s99
; GFX9-NEXT:    v_mov_b32_e32 v19, s96
; GFX9-NEXT:    v_mov_b32_e32 v13, s87
; GFX9-NEXT:    v_mov_b32_e32 v24, s82
; GFX9-NEXT:    v_mov_b32_e32 v14, s85
; GFX9-NEXT:    v_mov_b32_e32 v52, s84
; GFX9-NEXT:    v_mov_b32_e32 v11, s83
; GFX9-NEXT:    v_mov_b32_e32 v2, s4
; GFX9-NEXT:    v_mov_b32_e32 v25, s46
; GFX9-NEXT:    v_mov_b32_e32 v28, s56
; GFX9-NEXT:    v_mov_b32_e32 v34, s58
; GFX9-NEXT:    v_mov_b32_e32 v35, s60
; GFX9-NEXT:    v_mov_b32_e32 v36, s62
; GFX9-NEXT:    v_mov_b32_e32 v37, s72
; GFX9-NEXT:    v_mov_b32_e32 v38, s74
; GFX9-NEXT:    v_mov_b32_e32 v50, s76
; GFX9-NEXT:    v_mov_b32_e32 v26, s34
; GFX9-NEXT:    v_mov_b32_e32 v32, s36
; GFX9-NEXT:  .LBB91_5: ; %end
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v18, v18, v25, s4
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v50, v51, v50, s4
; GFX9-NEXT:    v_perm_b32 v23, v23, v26, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v50
; GFX9-NEXT:    v_perm_b32 v38, v49, v38, s4
; GFX9-NEXT:    v_perm_b32 v7, v7, v37, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_perm_b32 v36, v61, v36, s4
; GFX9-NEXT:    v_perm_b32 v6, v6, v35, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_perm_b32 v34, v56, v34, s4
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s4
; GFX9-NEXT:    v_perm_b32 v22, v22, v32, s4
; GFX9-NEXT:    v_readlane_b32 s99, v63, 35
; GFX9-NEXT:    v_readlane_b32 s98, v63, 34
; GFX9-NEXT:    v_readlane_b32 s97, v63, 33
; GFX9-NEXT:    v_readlane_b32 s96, v63, 32
; GFX9-NEXT:    v_readlane_b32 s87, v63, 31
; GFX9-NEXT:    v_readlane_b32 s86, v63, 30
; GFX9-NEXT:    v_readlane_b32 s85, v63, 29
; GFX9-NEXT:    v_readlane_b32 s84, v63, 28
; GFX9-NEXT:    v_readlane_b32 s83, v63, 27
; GFX9-NEXT:    v_readlane_b32 s82, v63, 26
; GFX9-NEXT:    v_readlane_b32 s81, v63, 25
; GFX9-NEXT:    v_readlane_b32 s80, v63, 24
; GFX9-NEXT:    v_readlane_b32 s71, v63, 23
; GFX9-NEXT:    v_readlane_b32 s70, v63, 22
; GFX9-NEXT:    v_readlane_b32 s69, v63, 21
; GFX9-NEXT:    v_readlane_b32 s68, v63, 20
; GFX9-NEXT:    v_readlane_b32 s67, v63, 19
; GFX9-NEXT:    v_readlane_b32 s66, v63, 18
; GFX9-NEXT:    v_readlane_b32 s65, v63, 17
; GFX9-NEXT:    v_readlane_b32 s64, v63, 16
; GFX9-NEXT:    v_readlane_b32 s55, v63, 15
; GFX9-NEXT:    v_readlane_b32 s54, v63, 14
; GFX9-NEXT:    v_readlane_b32 s53, v63, 13
; GFX9-NEXT:    v_readlane_b32 s52, v63, 12
; GFX9-NEXT:    v_readlane_b32 s51, v63, 11
; GFX9-NEXT:    v_readlane_b32 s50, v63, 10
; GFX9-NEXT:    v_readlane_b32 s49, v63, 9
; GFX9-NEXT:    v_readlane_b32 s48, v63, 8
; GFX9-NEXT:    v_readlane_b32 s39, v63, 7
; GFX9-NEXT:    v_readlane_b32 s38, v63, 6
; GFX9-NEXT:    v_readlane_b32 s37, v63, 5
; GFX9-NEXT:    v_readlane_b32 s36, v63, 4
; GFX9-NEXT:    v_readlane_b32 s35, v63, 3
; GFX9-NEXT:    v_readlane_b32 s34, v63, 2
; GFX9-NEXT:    v_readlane_b32 s31, v63, 1
; GFX9-NEXT:    v_readlane_b32 s30, v63, 0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v25, v30, s4
; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v25, v25, v31, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v26, v48, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v26, 16, v38
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v25, v39, s4
; GFX9-NEXT:    v_or_b32_e32 v25, v25, v26
; GFX9-NEXT:    buffer_store_dword v25, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v25, v17, s4
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v25, v25, v33, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX9-NEXT:    v_or_b32_e32 v17, v17, v25
; GFX9-NEXT:    buffer_store_dword v17, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v17, v17, v60, s4
; GFX9-NEXT:    v_or_b32_e32 v7, v17, v7
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v7, v7, v53, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v17, v5, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_or_b32_e32 v5, v7, v5
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v36
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v54, s4
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v5, v5, v55, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v7, v27, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v40, s4
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v5, v5, v41, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v10, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v34
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v42, s4
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v5, v5, v43, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v59, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v44, s4
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v5, v5, v45, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v58, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_perm_b32 v5, v8, v46, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v18
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v6, v9, v57, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v47, s4
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v5, v1, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v22
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v5
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v20, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v23
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v3, s4
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v1, v4, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v21, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v52, v2, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v11, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v1, v12, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v16, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v19, v2, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v1, v1, v14, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v24, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v15, s4
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v2, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v128i8_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-TRUE16-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s30, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s96, 0
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s31, 1
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s97, 1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v10
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s34, 2
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s98, 2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s35, 3
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s99, 3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s36, 4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s100, 4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s37, 5
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s101, 5
; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-TRUE16-NEXT:    s_clause 0x11 ; 72-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s38, 6
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s102, 6
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s39, 7
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s103, 7
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s48, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v75, s104, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s49, 9
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s50, 10
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s51, 11
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s52, 12
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s53, 13
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s54, 14
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s55, 15
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s64, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s65, 17
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s66, 18
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s67, 19
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s68, 20
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s69, 21
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s70, 22
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s71, 23
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s80, 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s81, 25
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s82, 26
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s83, 27
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s84, 28
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s85, 29
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s86, 30
; GFX11-TRUE16-NEXT:    v_writelane_b32 v74, s87, 31
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB91_3
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s83, s5, 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 10
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s97, s5, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s86, s5, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s96, s4, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 0
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s27, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s87, s4, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s71, s7, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s98, s7, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 11
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s82, s7, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s85, s6, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s84, s6, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s26, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s67, s9, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s99, s9, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s70, s9, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 9
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s81, s8, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s80, s8, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s55, s11, 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 14
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s100, s11, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s66, s11, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s69, s10, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 1
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s25, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s68, s10, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s51, s13, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s101, s13, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 15
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s24, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s54, s13, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s65, s12, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s64, s12, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 12
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s24, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s39, s15, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s102, s15, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s50, s15, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 13
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s53, s14, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s52, s14, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s35, s41, 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 18
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s103, s41, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s38, s41, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s49, s40, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 2
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s23, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s48, s40, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 vcc_hi, s29, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s104, s29, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 19
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s22, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s34, s29, 8
; GFX11-TRUE16-NEXT:    s_lshr_b32 s37, s28, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s36, s28, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 16
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s22, 8
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 17
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 22
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 16
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 3
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s21, 8
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 23
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s20, 16
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 20
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s20, 8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 21
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 26
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 4
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s19, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 27
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 24
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s18, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 25
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 30
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 5
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s17, 8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 31
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 28
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s16, 8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 29
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 24
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 2
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 6
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s3, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 3
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s2, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 0
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s2, 8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 1
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 6
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v76, s42, 7
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s1, 8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 7
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 16
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 4
; GFX11-TRUE16-NEXT:    s_lshr_b32 s42, s0, 8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    v_writelane_b32 v77, s42, 5
; GFX11-TRUE16-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB91_4
; GFX11-TRUE16-NEXT:  .LBB91_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_lshl_b32 s42, s29, 16
; GFX11-TRUE16-NEXT:    s_and_b32 s29, s29, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s42
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s29
; GFX11-TRUE16-NEXT:    s_and_b32 s42, s28, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s28, s28, 16
; GFX11-TRUE16-NEXT:    s_and_b32 s29, s41, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s42
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s29
; GFX11-TRUE16-NEXT:    s_lshl_b32 s41, s41, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v4, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s28, s40, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s29, s40, 16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v11, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v10, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v5, v8 :: v_dual_add_nc_u32 v5, 0x7fff, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s41
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v2
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v52.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v5, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v6, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s28
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-TRUE16-NEXT:    s_and_b32 s28, s15, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s15, s15, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v6, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v4, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    s_and_b32 s15, s14, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v3.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v10
; GFX11-TRUE16-NEXT:    s_lshl_b32 s14, s14, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX11-TRUE16-NEXT:    s_and_b32 s14, s13, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v6, v7
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v3, v11 :: v_dual_add_nc_u32 v6, v10, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s13, s13, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v3
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v8, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v9.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v10, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s13
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v5, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    s_and_b32 s13, s12, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s12, s12, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX11-TRUE16-NEXT:    s_and_b32 s12, s11, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    s_lshl_b32 s11, s11, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v8, v9
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v5, v13 :: v_dual_add_nc_u32 v8, v12, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v8, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v10, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v11.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v12, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s11
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v7, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    s_and_b32 s11, s10, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s10, s10, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v11
; GFX11-TRUE16-NEXT:    s_and_b32 s10, s9, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    s_lshl_b32 s9, s9, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v10, v11
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v7, v15 :: v_dual_add_nc_u32 v10, v14, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v54.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v10, v14, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v12, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v13.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v14, v16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v9, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    s_and_b32 s9, s8, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s8, s8, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v13
; GFX11-TRUE16-NEXT:    s_and_b32 s8, s7, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    s_lshl_b32 s7, s7, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v12, v13
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v9, v17 :: v_dual_add_nc_u32 v12, v16, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v66.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v12, v16, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, v14, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v15.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v16, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s7
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-TRUE16-NEXT:    s_and_b32 s7, s6, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s6, s6, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v15
; GFX11-TRUE16-NEXT:    s_and_b32 s6, s5, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v14, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    s_lshl_b32 s5, s5, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, v14, v15
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v11, v19 :: v_dual_add_nc_u32 v14, v18, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v64.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v17.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v16, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v18, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_and_b32 s5, s4, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v15
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v18, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s4, s4, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v20, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v15, v17
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    s_and_b32 s4, s1, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s1, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v13.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v21
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v20, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v22, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v21
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s0, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v17, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v23, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s3, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s3, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v13.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v17
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v25 :: v_dual_add_nc_u32 v17, v19, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v20, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s2, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s2, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v17, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v24, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s17, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v17
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v25 :: v_dual_add_nc_u32 v17, v19, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v20, v24
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.h, v13.l
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v23
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s16, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v17, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v24, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v20, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s19, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v20, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.h, v13.l
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s19, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v17
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v25 :: v_dual_add_nc_u32 v17, v19, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, v20, v24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v23
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s26, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s18, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v17, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v19, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v28, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.h, v13.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v14, v14, v19 :: v_dual_add_nc_u32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s21, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v29, v28, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v14
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v24 :: v_dual_add_nc_u32 v14, v18, v23
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v19
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v17, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, v29, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v13, v17, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v17, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v25, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v24, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v17, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s20, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x7fff, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s23, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v85.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v14
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v24 :: v_dual_add_nc_u32 v14, v18, v23
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, v18, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v23
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v18, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v18, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v26, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v25 :: v_dual_add_nc_u32 v18, v18, v24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s22, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v17
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v17, v19, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s25, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v17
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v18, v25 :: v_dual_add_nc_u32 v17, v19, v23
; GFX11-TRUE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v25, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x7fff, v17
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v18, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, v19, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v24, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v97.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v17, v18 :: v_dual_add_nc_u32 v18, 0x7fff, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v19, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v87.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, v118.l
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v18, v19 :: v_dual_add_nc_u32 v19, v23, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, v20, v25
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v18
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s24, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.h, v131.l
; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v19, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v176.l, v129.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, v26, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v20, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v19, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s27, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v18, v23, v24 :: v_dual_add_nc_u32 v23, v25, v19
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s27, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v26, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, v24, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v23, v25, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v26, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v27, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v27
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, v23, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v19
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v24, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v26
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x7fff, v23
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, v25, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 16, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v23, v23, v24, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x7fff, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v150.h, v147.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v176.h, v14.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v23
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v29, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v69.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v80.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v150.l, v144.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 16, v24
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v25, v25, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v70.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v82.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v86.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.l, v145.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v84.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v179.l, v116.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v179.h, v128.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, v119.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v149.h, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, v133.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.h, v135.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v161.l, v146.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v161.h, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.h, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v55.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v53.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v65.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v68.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v67.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v71.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v83.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v81.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v96.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v99.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v98.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v100.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v103.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v101.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v102.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.h, v114.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v112.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.l, v113.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v51.h, v117.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v115.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, v130.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.h, v134.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v163.l, v132.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v163.h, v17.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[26:27], 24, v[149:150]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[32:33], 24, v[176:177]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[17:18], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[161:162]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[178:179]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[18:19], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[23:24], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v179
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v179
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[163:164]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[50:51]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[36:37]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[30:31]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[13:14], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[24:25], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[1:2], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 24, v150
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 8, v150
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v149
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 8, v149
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 24, v162
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v162
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 16, v161
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v161
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 24, v164
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 8, v164
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v163
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v163
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v177
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v177
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v176
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v176
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v178
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v178
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v47, 24, v51
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v51
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v50
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v50
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v60, 24, v37
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v37
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v36
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v62, 24, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v63, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v176, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
; GFX11-TRUE16-NEXT:    s_branch .LBB91_5
; GFX11-TRUE16-NEXT:  .LBB91_3:
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr42
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr42
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr42
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr44
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr46
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr56
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr58
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr60
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr36
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr37
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr34
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr104
; GFX11-TRUE16-NEXT:    ; implicit-def: $vcc_hi
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr48
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr49
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr38
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr103
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr35
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr52
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr53
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr50
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr102
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr39
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr64
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr65
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr54
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr101
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr51
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr68
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr69
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr66
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr100
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr55
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr80
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr81
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr70
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr99
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr67
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr84
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr85
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr82
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr98
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr71
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr87
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr96
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr86
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr97
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr83
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr72
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr62
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr30
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr94
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr92
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr90
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr88
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr78
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr76
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr74
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-TRUE16-NEXT:    ; kill: killed $sgpr43
; GFX11-TRUE16-NEXT:    s_branch .LBB91_2
; GFX11-TRUE16-NEXT:  .LBB91_4:
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v112, s2 :: v_dual_mov_b32 v101, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 0
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, s41 :: v_dual_mov_b32 v145, s26
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v144, s27
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v146, s24 :: v_dual_mov_b32 v133, s25
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v147, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v52, s29
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v132, s22 :: v_dual_mov_b32 v129, s20
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v130, s23 :: v_dual_mov_b32 v119, s18
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v135, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 2
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v118, s21 :: v_dual_mov_b32 v115, s16
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v116, s19 :: v_dual_mov_b32 v113, s17
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v134, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 3
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v102, s3 :: v_dual_mov_b32 v87, s6
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v100, s1 :: v_dual_mov_b32 v85, s7
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v131, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 4
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v98, s4 :: v_dual_mov_b32 v81, s10
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v96, s5 :: v_dual_mov_b32 v71, s11
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v128, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 5
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v84, s8 :: v_dual_mov_b32 v69, s13
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v82, s9 :: v_dual_mov_b32 v67, s14
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v117, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 6
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v70, s12 :: v_dual_mov_b32 v65, s15
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v64, s40 :: v_dual_mov_b32 v53, s28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v114, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 7
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v99, s97 :: v_dual_mov_b32 v86, s99
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v97, s98 :: v_dual_mov_b32 v80, s101
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v103, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v83, s100 :: v_dual_mov_b32 v68, s102
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v66, s103 :: v_dual_mov_b32 v55, s104
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v148, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 9
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v25, s83 :: v_dual_mov_b32 v36, s55
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v29, s85 :: v_dual_mov_b32 v10, s66
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v49, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 10
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v15, s84 :: v_dual_mov_b32 v50, s51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v21, s71 :: v_dual_mov_b32 v8, s54
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v39, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 11
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v31, s81 :: v_dual_mov_b32 v178, s53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v11, s80 :: v_dual_mov_b32 v176, s39
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v35, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 12
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, s69 :: v_dual_mov_b32 v6, s50
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v9, s68 :: v_dual_mov_b32 v40, s48
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v160, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 13
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v51, s65 :: v_dual_mov_b32 v180, s35
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v7, s64 :: v_dual_mov_b32 v46, s37
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v150, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 14
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v5, s52 :: v_dual_mov_b32 v44, s36
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v41, s49 :: v_dual_mov_b32 v4, s34
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v151, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 15
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v179, s38 :: v_dual_mov_b32 v26, s62
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v43, vcc_hi :: v_dual_mov_b32 v28, s60
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v149, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 16
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v27, s72 :: v_dual_mov_b32 v32, s58
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, s56 :: v_dual_mov_b32 v34, s46
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v164, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 17
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v38, s44 :: v_dual_mov_b32 v13, s74
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v48, s42 :: v_dual_mov_b32 v17, s76
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v163, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 18
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s78 :: v_dual_mov_b32 v19, s88
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v23, s90 :: v_dual_mov_b32 v24, s92
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v162, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 19
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v3, s94
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v1, s30 :: v_dual_mov_b32 v22, s96
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s87 :: v_dual_mov_b32 v161, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 20
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v14, s86
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v16, s82
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v30, s67
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s70 :: v_dual_mov_b32 v177, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 21
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v166, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v167, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 23
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v165, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v42, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 25
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v181, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 26
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v183, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 27
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v182, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v57, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 29
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v56, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 30
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v47, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v76, 31
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v45, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v61, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v58, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 2
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v60, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 3
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v59, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 4
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v72, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 5
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v73, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v62, s0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s0, v77, 7
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v63, s0
; GFX11-TRUE16-NEXT:  .LBB91_5: ; %end
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v72, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v48, v101, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v101, v103, v62, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_perm_b32 v103, v100, v63, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v61, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v100, v114, v60, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v112, v112, v58, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v101, 16, v101
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; GFX11-TRUE16-NEXT:    v_perm_b32 v114, v102, v59, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v58, 16, v100
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v100, v48, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v57, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v102, v112, v38
; GFX11-TRUE16-NEXT:    v_perm_b32 v34, v115, v56, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v42, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v112, v128, v183, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v117, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v101, v103, v101
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v103, v114, v58
; GFX11-TRUE16-NEXT:    v_perm_b32 v114, v119, v181, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; GFX11-TRUE16-NEXT:    v_perm_b32 v115, v116, v182, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v116, 16, v112
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v112, v34, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v177, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v48, v113, v45, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v114, v114, v33
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v131, v167, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v129, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v164, v28, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v113, v48, v38
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v115, v115, v116
; GFX11-TRUE16-NEXT:    v_perm_b32 v34, v118, v165, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v132, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[100:103], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[112:115], off offset:16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v100, v33, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v134, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v160, v27, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v101, v34, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v102, v38, v28
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v130, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v32, v135, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v148, v26, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v33, v146, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v178, v24, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v37, v19, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v34, v133, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v38, 16, v32
; GFX11-TRUE16-NEXT:    v_perm_b32 v48, v145, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v103, v28, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v147, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, v33, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v55, v43, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v67, v5, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v51, v23, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v81, v9, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v31, v18, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, v34, v38
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, v48, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v144, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v52, v4, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v70, v7, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, v5, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v80, v50, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v84, v11, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, v9, v19
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v86, v30, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v46, v1, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v41, v3, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, v26, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v66, v180, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, v4, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v68, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, v7, v23
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v83, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v69, v8, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, v11, v18
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v29, v17, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v97, v21, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v22, v13, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v99, v25, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v53, v44, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v38, v64, v40, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v54, v179, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v65, v6, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v71, v10, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v82, v12, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v8, v23
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v87, v15, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v85, v16, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v98, v20, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v96, v14, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, v28, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, v38, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, v4, v26
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, v6, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, v12, v18
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, v15, v17
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v16, v19
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, v20, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v16, v23, v21
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[100:103], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[32:35], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:112
; GFX11-TRUE16-NEXT:    s_clause 0x11 ; 72-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-TRUE16-NEXT:    v_readlane_b32 s104, v75, 8
; GFX11-TRUE16-NEXT:    v_readlane_b32 s103, v75, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s102, v75, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s101, v75, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s100, v75, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s99, v75, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s98, v75, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s97, v75, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s96, v75, 0
; GFX11-TRUE16-NEXT:    v_readlane_b32 s87, v74, 31
; GFX11-TRUE16-NEXT:    v_readlane_b32 s86, v74, 30
; GFX11-TRUE16-NEXT:    v_readlane_b32 s85, v74, 29
; GFX11-TRUE16-NEXT:    v_readlane_b32 s84, v74, 28
; GFX11-TRUE16-NEXT:    v_readlane_b32 s83, v74, 27
; GFX11-TRUE16-NEXT:    v_readlane_b32 s82, v74, 26
; GFX11-TRUE16-NEXT:    v_readlane_b32 s81, v74, 25
; GFX11-TRUE16-NEXT:    v_readlane_b32 s80, v74, 24
; GFX11-TRUE16-NEXT:    v_readlane_b32 s71, v74, 23
; GFX11-TRUE16-NEXT:    v_readlane_b32 s70, v74, 22
; GFX11-TRUE16-NEXT:    v_readlane_b32 s69, v74, 21
; GFX11-TRUE16-NEXT:    v_readlane_b32 s68, v74, 20
; GFX11-TRUE16-NEXT:    v_readlane_b32 s67, v74, 19
; GFX11-TRUE16-NEXT:    v_readlane_b32 s66, v74, 18
; GFX11-TRUE16-NEXT:    v_readlane_b32 s65, v74, 17
; GFX11-TRUE16-NEXT:    v_readlane_b32 s64, v74, 16
; GFX11-TRUE16-NEXT:    v_readlane_b32 s55, v74, 15
; GFX11-TRUE16-NEXT:    v_readlane_b32 s54, v74, 14
; GFX11-TRUE16-NEXT:    v_readlane_b32 s53, v74, 13
; GFX11-TRUE16-NEXT:    v_readlane_b32 s52, v74, 12
; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v74, 11
; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v74, 10
; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v74, 9
; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v74, 8
; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v74, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v74, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v74, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v74, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s35, v74, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s34, v74, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s31, v74, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s30, v74, 0
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-TRUE16-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v128i8_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-FAKE16-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:84
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s30, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s96, 0
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s31, 1
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s97, 1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v10
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s34, 2
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s98, 2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s35, 3
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s99, 3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s36, 4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s100, 4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s37, 5
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s101, 5
; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-FAKE16-NEXT:    s_clause 0x11 ; 72-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s38, 6
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s102, 6
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s39, 7
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s103, 7
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s48, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v75, s104, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s49, 9
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s50, 10
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s51, 11
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s52, 12
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s53, 13
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s54, 14
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s55, 15
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s64, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s65, 17
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s66, 18
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s67, 19
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s68, 20
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s69, 21
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s70, 22
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s71, 23
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s80, 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s81, 25
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s82, 26
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s83, 27
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s84, 28
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s85, 29
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s86, 30
; GFX11-FAKE16-NEXT:    v_writelane_b32 v74, s87, 31
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB91_3
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s83, s5, 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 10
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s97, s5, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s86, s5, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s96, s4, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 0
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s27, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s87, s4, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s71, s7, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s98, s7, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 11
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s82, s7, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s85, s6, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s84, s6, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s26, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s67, s9, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s99, s9, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s70, s9, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 9
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s81, s8, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s80, s8, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s55, s11, 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 14
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s100, s11, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s66, s11, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s69, s10, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 1
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s25, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s68, s10, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s51, s13, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s101, s13, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 15
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s54, s13, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s65, s12, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s64, s12, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 12
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s24, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s39, s15, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s102, s15, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s50, s15, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 13
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s53, s14, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s52, s14, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s35, s41, 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 18
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s103, s41, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s38, s41, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s49, s40, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 2
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s23, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s48, s40, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 vcc_hi, s29, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s104, s29, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 19
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s34, s29, 8
; GFX11-FAKE16-NEXT:    s_lshr_b32 s37, s28, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s36, s28, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 16
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s22, 8
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 17
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 22
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 16
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 3
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s21, 8
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 23
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 16
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 20
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s20, 8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 21
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 26
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 4
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s19, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 27
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 24
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s18, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 25
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 30
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 5
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s17, 8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 31
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 28
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s16, 8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 29
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 24
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 2
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 6
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s3, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 3
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s2, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 0
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s2, 8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 1
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 6
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v76, s42, 7
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s1, 8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 7
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s0, 16
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 4
; GFX11-FAKE16-NEXT:    s_lshr_b32 s42, s0, 8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    v_writelane_b32 v77, s42, 5
; GFX11-FAKE16-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB91_4
; GFX11-FAKE16-NEXT:  .LBB91_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    s_lshl_b32 s42, s29, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s29, s29, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s42
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s29
; GFX11-FAKE16-NEXT:    s_and_b32 s42, s28, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s28, s28, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s42
; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v2
; GFX11-FAKE16-NEXT:    s_and_b32 s29, s41, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v6, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s41, s41, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s29
; GFX11-FAKE16-NEXT:    s_and_b32 s28, s40, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s29, s40, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v11, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v10, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v5, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s41
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v6, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v9 :: v_dual_add_nc_u32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s28, s15, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v49
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v1, v11 :: v_dual_add_nc_u32 v4, v6, v9
; GFX11-FAKE16-NEXT:    s_lshl_b32 s15, s15, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v50
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v48
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v8, v10
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s15
; GFX11-FAKE16-NEXT:    s_and_b32 s15, s14, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s14, s14, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v51, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s14
; GFX11-FAKE16-NEXT:    s_and_b32 s14, s13, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v5, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    s_lshl_b32 s13, s13, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v6, v7
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v4, v11 :: v_dual_add_nc_u32 v6, v10, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v54, 16, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v6, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v3, 16, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v10, v12
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v5, v13 :: v_dual_add_nc_u32 v6, v8, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v7, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v53
; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v8, v9
; GFX11-FAKE16-NEXT:    s_and_b32 s13, s12, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s12, s12, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v12, v10
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
; GFX11-FAKE16-NEXT:    s_and_b32 s12, s11, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX11-FAKE16-NEXT:    s_lshl_b32 s11, s11, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 16, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v6, v13, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v64, 16, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v55
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 16, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 8, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 24, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v5, 16, v11
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v12, v14
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v15 :: v_dual_add_nc_u32 v8, v10, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v7
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xffff, v65
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v8, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v68, 16, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v10, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v14, v12
; GFX11-FAKE16-NEXT:    s_and_b32 s11, s10, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v66
; GFX11-FAKE16-NEXT:    s_lshl_b32 s10, s10, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s10
; GFX11-FAKE16-NEXT:    s_and_b32 s10, s9, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-FAKE16-NEXT:    s_lshl_b32 s9, s9, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v15, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v16, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v14, v16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v9, v17 :: v_dual_add_nc_u32 v10, v12, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v67
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s9
; GFX11-FAKE16-NEXT:    s_and_b32 s9, s8, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s8, s8, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v10, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v71, 16, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s8
; GFX11-FAKE16-NEXT:    s_and_b32 s8, s7, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v69
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v12, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v17, 0x40c00000, s9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v9, 16, v15
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s8
; GFX11-FAKE16-NEXT:    s_lshl_b32 s7, s7, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v16, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v17, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v12, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v12
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v16, v18
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v11, v11, v19 :: v_dual_add_nc_u32 v12, v14, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v16, 0x40c00000, s7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v16, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s7, s6, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-FAKE16-NEXT:    s_lshl_b32 s6, s6, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v14, v15
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v17, v12, v19 :: v_dual_add_nc_u32 v14, v18, v16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
; GFX11-FAKE16-NEXT:    s_and_b32 s6, s5, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s5, s5, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v70
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v18, v20
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v13, v21 :: v_dual_add_nc_u32 v14, v16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v82, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v17
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff, v80
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v16, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v17
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, v20, v21
; GFX11-FAKE16-NEXT:    s_and_b32 s5, s4, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s4, s4, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x7fff, v16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v81
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v85, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v83
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v19, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v22
; GFX11-FAKE16-NEXT:    s_and_b32 s4, s1, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s1, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v13, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s4
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v20
; GFX11-FAKE16-NEXT:    s_and_b32 s1, s0, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s0, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v84
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v18, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v22, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v87, 16, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v86
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v18
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v15, v15, v24 :: v_dual_add_nc_u32 v18, v20, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v23
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s3, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s3, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v13, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 16, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v22
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s2, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s2, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v18, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v96
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v18, v19
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v23, v20
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v23, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v32, v99, 16, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v21, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v23, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v97
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v18
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v15, v15, v24 :: v_dual_add_nc_u32 v18, v20, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v21, v23
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v13, 16, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 16, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v18, 0x400000, v23
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v20, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s17, 16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v98
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v18, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v39, v102, 16, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v21, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 16, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v18, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v19 :: v_dual_add_nc_u32 v13, 0x7fff, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s16, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v100
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, v24, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v20, v27, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v23, vcc_lo
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v38, v14, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 16, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v20, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s19, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v19, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v101
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s21, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v22, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v19
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v22, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 16, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s18, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v20, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v112
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v179, v113, 16, v21
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v22, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v178, v13, 16, v23
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v15
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v21, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v103
; GFX11-FAKE16-NEXT:    v_bfe_u32 v22, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v15
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v181, v116, 16, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, v21, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v22, v15
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, v23, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x7fff, v20
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x7fff, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 24, v181
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v20, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 24, v179
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 16, v14
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v22, v27, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s20, 16
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff, v115
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v21, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v19, v14, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s23, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 16, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff, v114
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v19, v14
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v13, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v165, v118, 16, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v22, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v22, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, v23, v21
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v19, v24 :: v_dual_add_nc_u32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, v27, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v21
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x7fff, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, 0x400000, v22
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v21, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s22, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v21, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v19, v24, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v21
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v117
; GFX11-FAKE16-NEXT:    v_bfe_u32 v23, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v24, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s25, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v162, v129, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, 0x400000, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, v23, v18
; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v24, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v119
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v19, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v19, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v28, v19, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s24, 16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v21, v23 :: v_dual_add_nc_u32 v21, v27, v24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, v28, v19
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v29, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 16, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, v27, v29
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v30, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s27, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v23, 0x400000, v29
; GFX11-FAKE16-NEXT:    v_bfe_u32 v27, v30, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 16, v19
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v18, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, 0x400000, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v180, v13, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 16, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v18, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s27, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, v27, v30
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v27, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s26, 16
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v29, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x7fff, v21
; GFX11-FAKE16-NEXT:    v_bfe_u32 v33, v27, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v24, v18, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v35, v29, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v21, v28, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, v33, v27
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v34, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, v35, v29
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v28
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, v24, v18
; GFX11-FAKE16-NEXT:    v_bfe_u32 v30, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x7fff, v33
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v28, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x7fff, v24
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v18
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, v30, v34
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v33, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 16, v27
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x7fff, v30
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v24, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff, v128
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff, v134
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 16, v18
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v28, v30, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v130
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff, v131
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v167, v133, 16, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v166, v19, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v132
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v161, v15, 16, v28
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v146, v135, 16, v18
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v164, v14, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[180:181]
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v145, v21, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[18:19], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[27:28], 24, v[166:167]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[178:179]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[19:20], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[28:29], 24, v[161:162]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[38:39]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[13:14], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[20:21], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v180
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 8, v180
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v179
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[3:4], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[23:24], 24, v[145:146]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[29:30], 24, v[164:165]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[14:15], 24, v[16:17]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[21:22], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[1:2], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v146
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 8, v146
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 16, v145
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v145
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 24, v167
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 8, v167
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 16, v166
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v166
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v162
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v162
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v161
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v161
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v165
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v165
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 16, v164
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v164
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v178
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v178
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 24, v39
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 8, v39
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 16, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v38
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v15, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 24, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
; GFX11-FAKE16-NEXT:    s_branch .LBB91_5
; GFX11-FAKE16-NEXT:  .LBB91_3:
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vcc_hi
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr30
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $sgpr43
; GFX11-FAKE16-NEXT:    ; kill: killed $sgpr43
; GFX11-FAKE16-NEXT:    s_branch .LBB91_2
; GFX11-FAKE16-NEXT:  .LBB91_4:
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v100, s2 :: v_dual_mov_b32 v97, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 0
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v50, s41
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v132, s26 :: v_dual_mov_b32 v131, s27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v134, s24 :: v_dual_mov_b32 v117, s23
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v135, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v48, s29
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v128, s25 :: v_dual_mov_b32 v119, s20
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v130, s22 :: v_dual_mov_b32 v115, s18
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v133, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 2
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v114, s21 :: v_dual_mov_b32 v103, s19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v112, s16 :: v_dual_mov_b32 v101, s17
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v129, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 3
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v98, s3 :: v_dual_mov_b32 v83, s6
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v96, s1 :: v_dual_mov_b32 v81, s7
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v118, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 4
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v86, s4 :: v_dual_mov_b32 v69, s10
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, s5 :: v_dual_mov_b32 v67, s11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v116, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v80, s8 :: v_dual_mov_b32 v65, s13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, s9 :: v_dual_mov_b32 v55, s14
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v113, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 6
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, s12 :: v_dual_mov_b32 v53, s15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, s40 :: v_dual_mov_b32 v49, s28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v102, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v87, s97 :: v_dual_mov_b32 v82, s99
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, s98 :: v_dual_mov_b32 v68, s101
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v99, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 8
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v71, s100 :: v_dual_mov_b32 v64, s102
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, s103 :: v_dual_mov_b32 v51, s104
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v146, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v15, s86 :: v_dual_mov_b32 v16, s84
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v25, s71 :: v_dual_mov_b32 v32, s81
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v145, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 10
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v17, s82 :: v_dual_mov_b32 v12, s70
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v11, s80 :: v_dual_mov_b32 v38, s55
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v144, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 11
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v31, s67 :: v_dual_mov_b32 v10, s66
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v39, s69 :: v_dual_mov_b32 v164, s65
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v37, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 12
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v9, s68 :: v_dual_mov_b32 v8, s54
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v7, s64 :: v_dual_mov_b32 v178, s53
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v150, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v161, s51 :: v_dual_mov_b32 v176, s39
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v5, s52 :: v_dual_mov_b32 v6, s50
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v148, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 14
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v41, s49 :: v_dual_mov_b32 v40, s48
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v180, s35 :: v_dual_mov_b32 v179, s38
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v149, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v46, s37 :: v_dual_mov_b32 v43, vcc_hi
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v44, s36 :: v_dual_mov_b32 v23, s62
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v147, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 16
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s34 :: v_dual_mov_b32 v27, s72
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s60 :: v_dual_mov_b32 v29, s58
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v163, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v33, s56 :: v_dual_mov_b32 v34, s46
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v35, s44 :: v_dual_mov_b32 v36, s42
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v162, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 18
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v13, s74 :: v_dual_mov_b32 v14, s76
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s78 :: v_dual_mov_b32 v19, s88
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v160, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s90 :: v_dual_mov_b32 v21, s92
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v3, s94
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v1, s30
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v151, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 20
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v26, s96
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v22, s87
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v24, s83
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s85 :: v_dual_mov_b32 v177, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 21
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v166, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v167, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 23
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v165, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 24
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v42, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 25
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v181, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 26
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v183, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 27
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v182, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v57, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 29
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v56, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 30
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v47, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v76, 31
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v45, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v61, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v58, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v60, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 3
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v59, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v72, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 5
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v73, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v62, s0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s0, v77, 7
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v63, s0
; GFX11-FAKE16-NEXT:  .LBB91_5: ; %end
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v72, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v97, v73, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v97, v99, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v99, v96, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v61, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v96, v102, v60, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v100, v100, v58, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v97, 16, v97
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v102, v98, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v58, 16, v96
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v96, v36, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v57, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v98, v100, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v34, v112, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v113, v47, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v33, v42, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v100, v116, v183, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v97, v99, v97
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v99, v102, v58
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v101, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v102, 16, v33
; GFX11-FAKE16-NEXT:    v_perm_b32 v103, v103, v182, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v100, 16, v100
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v34, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v177, v29, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v101, v115, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v36, v35
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v103, v100
; GFX11-FAKE16-NEXT:    v_perm_b32 v100, v119, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v163, v28, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v101, v102
; GFX11-FAKE16-NEXT:    v_perm_b32 v102, v130, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[96:99], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[33:36], off offset:16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v100, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v129, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v150, v27, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v146, v23, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v102, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v117, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v178, v21, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v118, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v96, v134, v148, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v98, v132, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v28, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v135, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v55, v5, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v164, v20, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v39, v19, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v101, v114, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v96, v96, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v51, v43, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v98, v98, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v131, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v66, v7, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, v5, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v68, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v69, v9, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v32, v18, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v101, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v133, v149, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v48, v4, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v99, v23, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v54, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, v7, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v65, v8, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v71, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v80, v11, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v82, v31, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v97, v128, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v46, v1, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v41, v3, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v64, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v50, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v67, v10, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, v8, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v30, v14, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v85, v25, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v26, v13, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v87, v24, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v97, v97, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v49, v44, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v52, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v53, v6, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v70, v12, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v4, v23
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v83, v16, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v81, v17, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v86, v22, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v22, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v84, v15, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v28, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v29, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, v6, v27
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v12, v18
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v16, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v17, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v21, v22
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v23, v20
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[33:36], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[96:99], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x11 ; 72-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-FAKE16-NEXT:    v_readlane_b32 s104, v75, 8
; GFX11-FAKE16-NEXT:    v_readlane_b32 s103, v75, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s102, v75, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s101, v75, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s100, v75, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s99, v75, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s98, v75, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s97, v75, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s96, v75, 0
; GFX11-FAKE16-NEXT:    v_readlane_b32 s87, v74, 31
; GFX11-FAKE16-NEXT:    v_readlane_b32 s86, v74, 30
; GFX11-FAKE16-NEXT:    v_readlane_b32 s85, v74, 29
; GFX11-FAKE16-NEXT:    v_readlane_b32 s84, v74, 28
; GFX11-FAKE16-NEXT:    v_readlane_b32 s83, v74, 27
; GFX11-FAKE16-NEXT:    v_readlane_b32 s82, v74, 26
; GFX11-FAKE16-NEXT:    v_readlane_b32 s81, v74, 25
; GFX11-FAKE16-NEXT:    v_readlane_b32 s80, v74, 24
; GFX11-FAKE16-NEXT:    v_readlane_b32 s71, v74, 23
; GFX11-FAKE16-NEXT:    v_readlane_b32 s70, v74, 22
; GFX11-FAKE16-NEXT:    v_readlane_b32 s69, v74, 21
; GFX11-FAKE16-NEXT:    v_readlane_b32 s68, v74, 20
; GFX11-FAKE16-NEXT:    v_readlane_b32 s67, v74, 19
; GFX11-FAKE16-NEXT:    v_readlane_b32 s66, v74, 18
; GFX11-FAKE16-NEXT:    v_readlane_b32 s65, v74, 17
; GFX11-FAKE16-NEXT:    v_readlane_b32 s64, v74, 16
; GFX11-FAKE16-NEXT:    v_readlane_b32 s55, v74, 15
; GFX11-FAKE16-NEXT:    v_readlane_b32 s54, v74, 14
; GFX11-FAKE16-NEXT:    v_readlane_b32 s53, v74, 13
; GFX11-FAKE16-NEXT:    v_readlane_b32 s52, v74, 12
; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v74, 11
; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v74, 10
; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v74, 9
; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v74, 8
; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v74, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v74, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v74, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v74, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s35, v74, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s34, v74, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s31, v74, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s30, v74, 0
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-FAKE16-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:84
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define <64 x half> @bitcast_v128i8_to_v64f16(<128 x i8> %a, i32 %b) {
; SI-LABEL: bitcast_v128i8_to_v64f16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v59, v24
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v47, v16
; SI-NEXT:    v_mov_b32_e32 v61, v14
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v60, v6
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:388
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92
; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v25
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:4
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v7
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v15
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v11
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v21
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v23
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v19
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v29
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v27
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v43, v8
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v26
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:116
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:36
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:108
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:28
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v6
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:100
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:136
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(5) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v5
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v30
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:168
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v10
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v32
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v33
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v34
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:188
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:196
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v2
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v11
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:232
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:228
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v12
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:268
; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v2
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:316
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:264
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:260
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:300
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v13
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:336
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:332
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:296
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:292
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:328
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 24, v3
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v14
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v2
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:352
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:376
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v0
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:384
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:380
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:368
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:360
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:356
; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v18
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 24, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v20
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:68
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:60
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:20
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:44
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB92_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v51, 0xff, v47
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v48, 0xff, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_or_b32_e32 v0, v57, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v58
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v62, v3
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v5, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v60
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_mov_b32_e32 v60, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v5, v1, v5
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v9, v1, v9
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v10, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v61
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v10, v1, v10
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v1, v12
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v13, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v1, v14
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v1, v14
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v15, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v1, v16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v1, v16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v19, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v49, v1, v20
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v1, v20
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v21, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v50, v1, v22
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v1, v22
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v23, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v52, v1, v24
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v1, v24
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v27, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v53, v1, v28
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v1, v28
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v29, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v54, v1, v30
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v1, v30
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v33, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v55, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v34, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v35, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v41, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v36, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v37, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v42, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v38, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v39, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v44, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v45, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v63, v31, v63
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v26, v26, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v7, v7, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v32, v31, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v46, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v46, v56, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v31, v31, v40
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v59
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v31, v9
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v48, v48, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v51, v51, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v56, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v57, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v58, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v62, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v6, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v40, v1
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v60
; SI-NEXT:    v_or_b32_e32 v40, v40, v43
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v60, v43, v47
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v47, v43, v47
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v61, v43, v59
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v43, v43, v59
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v59, 0xff, v59
; SI-NEXT:    v_or_b32_e32 v59, v59, v8
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v5
; SI-NEXT:    v_alignbit_b32 v9, v8, v9, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v9, v10
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v48
; SI-NEXT:    v_or_b32_e32 v48, v9, v12
; SI-NEXT:    v_alignbit_b32 v9, v11, v12, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v12, v9, v17
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v51
; SI-NEXT:    v_or_b32_e32 v9, v9, v14
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v9, v12, v14, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v51, v9, v25
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v56
; SI-NEXT:    v_or_b32_e32 v15, v9, v16
; SI-NEXT:    v_alignbit_b32 v9, v51, v16, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v9, v9, v49
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v57
; SI-NEXT:    v_or_b32_e32 v16, v13, v20
; SI-NEXT:    v_alignbit_b32 v13, v9, v20, 16
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v31, v13, v50
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v58
; SI-NEXT:    v_or_b32_e32 v13, v13, v22
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v13, v31, v22, 16
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v62
; SI-NEXT:    v_or_b32_e32 v13, v13, v52
; SI-NEXT:    v_or_b32_e32 v14, v14, v24
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v14, v13, v24, 16
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v24, v14, v53
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v14, v28
; SI-NEXT:    v_alignbit_b32 v14, v24, v28, 16
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v21, v14, v54
; SI-NEXT:    v_or_b32_e32 v22, v6, v30
; SI-NEXT:    v_alignbit_b32 v6, v21, v30, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v33
; SI-NEXT:    v_or_b32_e32 v19, v6, v55
; SI-NEXT:    v_or_b32_e32 v29, v1, v34
; SI-NEXT:    v_alignbit_b32 v1, v19, v34, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v35
; SI-NEXT:    v_or_b32_e32 v23, v1, v41
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v40
; SI-NEXT:    v_or_b32_e32 v20, v1, v36
; SI-NEXT:    v_alignbit_b32 v1, v23, v36, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v37
; SI-NEXT:    v_or_b32_e32 v30, v1, v42
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v60
; SI-NEXT:    v_or_b32_e32 v27, v1, v38
; SI-NEXT:    v_alignbit_b32 v1, v30, v38, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v39
; SI-NEXT:    v_or_b32_e32 v28, v1, v44
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v47
; SI-NEXT:    v_or_b32_e32 v34, v1, v45
; SI-NEXT:    v_alignbit_b32 v1, v28, v45, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
; SI-NEXT:    v_or_b32_e32 v33, v1, v26
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
; SI-NEXT:    v_or_b32_e32 v36, v1, v7
; SI-NEXT:    v_alignbit_b32 v1, v33, v7, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v32
; SI-NEXT:    v_or_b32_e32 v35, v1, v4
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v43
; SI-NEXT:    v_or_b32_e32 v38, v1, v46
; SI-NEXT:    v_alignbit_b32 v1, v35, v46, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v37, v1, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v39, v1, v3
; SI-NEXT:    v_alignbit_b32 v1, v37, v3, 16
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v50
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v52
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v54
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v42
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v4
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:  .LBB92_2: ; %Flow
; SI-NEXT:    s_or_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
; SI-NEXT:    s_xor_b64 exec, exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB92_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_or_b32_e32 v2, v58, v2
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_movk_i32 s6, 0x300
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_or_b32_e32 v0, v57, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v62, v3
; SI-NEXT:    s_mov_b32 s7, 0x3000000
; SI-NEXT:    v_add_i32_e32 v37, vcc, s7, v0
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v37
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v26, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v5, v1, v5
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, 0x300, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v3, v3, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v39, vcc, s7, v3
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v56, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v38, vcc, s7, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v46, v4
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v35, vcc, s7, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v35
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v7, v5
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v36, vcc, s7, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v63, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v33, vcc, s7, v5
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v33
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v34, vcc, s7, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v7
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v8
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v32
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_add_i32_e32 v30, vcc, s7, v9
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v11
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v12
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v17
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v18
; SI-NEXT:    v_mov_b32_e32 v18, v5
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v25, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v26, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v26
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v31, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v15
; SI-NEXT:    v_add_i32_e32 v31, vcc, s7, v31
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v32, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v32
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v48, v20, v19
; SI-NEXT:    v_or_b32_e32 v0, v48, v0
; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v25
; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v10
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v59
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v51, vcc, s7, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v51
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v48, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v40, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v60
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v0
; SI-NEXT:    v_alignbit_b32 v0, v11, v48, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v12, v6, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v51, v15, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v9, v16, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v31, v3, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v13, v4, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v24, v5, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v21, v22, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v19, v29, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v23, v20, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v30, v27, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v28, v34, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v33, v36, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v35, v38, 16
; SI-NEXT:    v_alignbit_b32 v1, v8, v7, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v37, v39, 16
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
; SI-NEXT:  .LBB92_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v6
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v41
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v45
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v48
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v8, v8, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v17
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v49
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v52
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v25
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v29
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v54
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v55
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v18, v18, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v53
; SI-NEXT:    v_or_b32_e32 v19, v19, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v34
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v44
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v36
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v33
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v38
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v35
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v39
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v37
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v64f16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:4
; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:376
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:120
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB92_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s6, 0xc0c0004
; VI-NEXT:    v_perm_b32 v10, v37, v36, s6
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v11, v59, v34, s6
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_perm_b32 v12, v57, v32, s6
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v47, v46, s6
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v43, v42, s6
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v55, v54, s6
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v16, v51, v50, s6
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v1, v0, s6
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s6
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v9, s6
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_perm_b32 v9, v39, v62, s6
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_perm_b32 v8, v38, v61, s6
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_perm_b32 v9, v35, v60, s6
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_perm_b32 v10, v33, v58, s6
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_perm_b32 v11, v63, v56, s6
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v45, v44, s6
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_perm_b32 v13, v41, v40, s6
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_perm_b32 v14, v53, v52, s6
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v49, v15, s6
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v17, v18, v19
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v20, v21, s6
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v22, s6
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v27, v28, s6
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v29, v30, s6
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v30, v30, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v31, v31, v48, s6
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v32, v33, v32, s6
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:  .LBB92_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB92_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v18, 0x300
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v61
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v60
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v58
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v56
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v44
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v40
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v52
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    s_waitcnt vmcnt(11)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    s_waitcnt vmcnt(10)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    s_waitcnt vmcnt(9)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    s_waitcnt vmcnt(8)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    s_waitcnt vmcnt(7)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    s_waitcnt vmcnt(6)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v9
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v1, 0x300, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v3
; VI-NEXT:    v_or_b32_e32 v0, v2, v0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v1, v1, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
; VI-NEXT:    v_or_b32_e32 v2, v3, v2
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
; VI-NEXT:    v_or_b32_e32 v3, v4, v3
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
; VI-NEXT:    v_or_b32_e32 v4, v5, v4
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v5, v5, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
; VI-NEXT:    v_or_b32_e32 v5, v6, v5
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
; VI-NEXT:    v_or_b32_e32 v6, v7, v6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
; VI-NEXT:    v_or_b32_e32 v7, v8, v7
; VI-NEXT:    v_add_u16_e32 v8, 3, v39
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v62
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 3, v38
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
; VI-NEXT:    v_or_b32_e32 v8, v9, v8
; VI-NEXT:    v_add_u16_e32 v9, 3, v37
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v36
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 3, v35
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
; VI-NEXT:    v_or_b32_e32 v9, v10, v9
; VI-NEXT:    v_add_u16_e32 v10, 3, v59
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v34
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 3, v33
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
; VI-NEXT:    v_or_b32_e32 v10, v11, v10
; VI-NEXT:    v_add_u16_e32 v11, 3, v57
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v32
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 3, v63
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
; VI-NEXT:    v_or_b32_e32 v11, v12, v11
; VI-NEXT:    v_add_u16_e32 v12, 3, v47
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v46
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 3, v45
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    v_add_u16_e32 v13, 3, v43
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v42
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 3, v41
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_add_u16_e32 v14, 3, v55
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v54
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v15, 3, v53
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_add_u16_e32 v15, 3, v51
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v50
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v16, 3, v49
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v21, v16, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v19, 0x300, v20
; VI-NEXT:    v_or_b32_e32 v16, v19, v16
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v17, v17, v21
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v48
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v31, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v31, 0x300, v31
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v30, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v33, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v29, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v35, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v28, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v35
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v37, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v27, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v37
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v38, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v26, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v38
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v39, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v39, v39, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v25, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v39
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v48, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v48, v48, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v24, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v48
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v49, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v49, v49, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v23, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v49
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v50, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v50, v50, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v22, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v50
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v51, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v51, v51, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    v_or_b32_sdwa v21, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v51
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    v_or_b32_sdwa v52, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v36, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v36, 0x300, v36
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v53, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v54, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v55, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v32, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v34, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v20, v52, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v19, v53, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v18, v55, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v36, v20
; VI-NEXT:    v_or_b32_e32 v30, v30, v34
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v52, 3, v52
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v53
; VI-NEXT:    v_or_b32_sdwa v52, v53, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v52, 0x300, v52
; VI-NEXT:    v_or_b32_e32 v18, v52, v18
; VI-NEXT:    v_add_u16_e32 v52, 0x300, v54
; VI-NEXT:    v_or_b32_e32 v19, v52, v19
; VI-NEXT:  .LBB92_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v64f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:264
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:272
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:288
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:296
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:304
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:320
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
; GFX9-NEXT:    s_waitcnt vmcnt(27)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-NEXT:    s_waitcnt vmcnt(26)
; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v1
; GFX9-NEXT:    s_waitcnt vmcnt(25)
; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(24)
; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(23)
; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v4
; GFX9-NEXT:    s_waitcnt vmcnt(22)
; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v6
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:100
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:92
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v7
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v9
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v11
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v13
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v15
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v17
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:352
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:348
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:344
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:360
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:384
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:356
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:364
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:380
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:340
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:12
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:372
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:376
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB92_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0xc0c0004
; GFX9-NEXT:    s_mov_b32 s7, 0x5040100
; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v10, v47, v45, s6
; GFX9-NEXT:    v_or_b32_sdwa v11, v40, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v12, v55, v54, s6
; GFX9-NEXT:    v_perm_b32 v17, v43, v42, s6
; GFX9-NEXT:    v_perm_b32 v18, v58, v57, s6
; GFX9-NEXT:    v_perm_b32 v19, v63, v46, s6
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_perm_b32 v1, v3, v2, s7
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s7
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s7
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s7
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s7
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s7
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v33, s6
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s7
; GFX9-NEXT:    v_or_b32_sdwa v8, v60, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s7
; GFX9-NEXT:    v_or_b32_sdwa v9, v44, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v9, v9, v10, s7
; GFX9-NEXT:    v_or_b32_sdwa v10, v41, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s7
; GFX9-NEXT:    v_or_b32_sdwa v11, v52, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v11, v11, v12, s7
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v12, v50, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s7
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v14, v37, v36, s6
; GFX9-NEXT:    v_perm_b32 v13, v13, v14, s7
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s7
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_or_b32_sdwa v15, v62, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v35, s6
; GFX9-NEXT:    v_perm_b32 v15, v15, v16, s7
; GFX9-NEXT:    v_perm_b32 v16, v32, v56, s6
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s7
; GFX9-NEXT:    v_perm_b32 v17, v19, v18, s7
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s6
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s7
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s7
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v20, v20, v61, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s7
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s7
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s7
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s7
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s7
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s7
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s7
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s7
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s7
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s7
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s7
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v32, v33, v32, s6
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s7
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:  .LBB92_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB92_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v33
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v45
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v40
; GFX9-NEXT:    v_or_b32_sdwa v11, v38, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v54
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v46
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v56
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v42
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(19)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v1, 0x300, v1
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v3
; GFX9-NEXT:    v_perm_b32 v0, v2, v0, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v1, v1, v4, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v60
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
; GFX9-NEXT:    v_or_b32_sdwa v8, v53, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_or_b32_sdwa v9, v51, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v47
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v44
; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v41
; GFX9-NEXT:    v_or_b32_sdwa v10, v39, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v55
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v52
; GFX9-NEXT:    v_or_b32_sdwa v12, v34, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v50
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v48
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v37
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v36
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v14
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v14
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v35
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v62
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v58
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v57
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v63
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v32
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v43
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v16
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v18
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v19
; GFX9-NEXT:    v_perm_b32 v16, v18, v16, s6
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
; GFX9-NEXT:    v_perm_b32 v17, v17, v20, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v32, 0x300, v19
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v33, 0x300, v21
; GFX9-NEXT:    v_perm_b32 v30, v33, v30, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v34, 0x300, v23
; GFX9-NEXT:    v_perm_b32 v29, v34, v29, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v35, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v36, 0x300, v36
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v39
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v39, 0x300, v50
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v51
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v53
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v54, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v55
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v61
; GFX9-NEXT:    v_add_u16_e32 v50, 0x300, v40
; GFX9-NEXT:    v_perm_b32 v21, v50, v21, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v41
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v51, 0x300, v42
; GFX9-NEXT:    v_perm_b32 v20, v51, v20, s6
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v43
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v44, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v45, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v45
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v46, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v24
; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v26
; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v37
; GFX9-NEXT:    v_add_u16_e32 v37, 0x300, v38
; GFX9-NEXT:    v_add_u16_e32 v38, 0x300, v48
; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v49
; GFX9-NEXT:    v_add_u16_e32 v48, 0x300, v52
; GFX9-NEXT:    v_add_u16_e32 v49, 0x300, v54
; GFX9-NEXT:    v_add_u16_e32 v52, 0x300, v44
; GFX9-NEXT:    v_add_u16_e32 v53, 0x300, v46
; GFX9-NEXT:    v_perm_b32 v18, v53, v18, s6
; GFX9-NEXT:    v_perm_b32 v19, v52, v19, s6
; GFX9-NEXT:    v_perm_b32 v22, v49, v22, s6
; GFX9-NEXT:    v_perm_b32 v23, v48, v23, s6
; GFX9-NEXT:    v_perm_b32 v24, v39, v24, s6
; GFX9-NEXT:    v_perm_b32 v25, v38, v25, s6
; GFX9-NEXT:    v_perm_b32 v26, v37, v26, s6
; GFX9-NEXT:    v_perm_b32 v27, v36, v27, s6
; GFX9-NEXT:    v_perm_b32 v28, v35, v28, s6
; GFX9-NEXT:  .LBB92_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64f16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:392
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:384
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:380
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:376
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:372
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v31, off, s32 offset:368
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:364
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v70, off, s32 offset:360
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v39, off, s32 offset:356
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:352
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:348
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:344
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:340
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:336
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v35, off, s32 offset:332
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v69, off, s32 offset:328
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v38, off, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v68, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v36, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v67, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v66, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v32, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v33, off, s32 offset:260
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:132
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v65, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:388
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v52, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v53, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v54, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v55, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v64, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v71, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v51, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v50, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v48, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v49, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v37, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v34, off, s32 offset:12
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.l, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.l, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.l, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, v0.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v69.l, 8, v69.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v70.l, 8, v70.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v69.h, 8, v69.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v68.h, 8, v68.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.l, 8, v67.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v68.l, 8, v68.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.h, 8, v67.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.h, 8, v66.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.l, 8, v66.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(33)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v65.h, 8, v65.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(29)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v88
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.l, 8, v52.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.h, 8, v52.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(25)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v53.l, 8, v53.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(24)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.l, 8, v53.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v53.h, 8, v54.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(22)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.l, 8, v55.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.h, 8, v55.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(20)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.h, 8, v64.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v65.l, 8, v64.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(18)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v64.l, 8, v71.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v64.h, 8, v71.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.l, 8, v70.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v70.h, 8, v31.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.h, 8, v31.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB92_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v130, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v165, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v45, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v60, v41, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v37.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v160, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v48.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v3.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v177, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v13.h, 0xff, v51.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v40, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v32.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v7.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v58, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v9.l, v53.l
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v49.h
; GFX11-TRUE16-NEXT:    v_and_b16 v11.l, 0xff, v38.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v11.h, v55.l
; GFX11-TRUE16-NEXT:    v_and_b16 v12.h, 0xff, v50.h
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v49.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v13.h, v65.l
; GFX11-TRUE16-NEXT:    v_and_b16 v15.l, 0xff, v51.h
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v50.l
; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v14.h, v65.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v63, v56, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v98, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v10.h, v54.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v53.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v77, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v102, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v12.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.h, v55.h
; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v13.l, v54.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v79, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v119, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v14.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v15.l, v64.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v15.h, v64.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v83, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v144, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v16.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v96, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v149, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v17.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v99, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v163, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v18.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v103, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v179, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v117, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v19.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v148, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v44, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v25.l, 0xff, v34.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v146, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v20.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v145, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v35.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v181, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v21.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v161, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v39.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v57, v182, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v22.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v43, v167, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v48.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v113, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v23.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v180, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v2.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v147, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v26.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v25.l, v67.h
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v36.h
; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v33.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v4.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v164, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v27.h, v68.h
; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v38.h
; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v35.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v62, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v29.h, v69.l
; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v39.h
; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v36.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v37.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.h, 0xff, v34.h
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v33.h
; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v32.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v30.l, v71.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v59, v183, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v26.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v26.h, v68.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v27.l, v67.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v72, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v28.l
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.h, v69.h
; GFX11-TRUE16-NEXT:    v_or_b16 v28.h, v29.l, v70.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v74, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v30.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v31.l, v71.l
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v31.h, v70.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v78, v75, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.l, v52.l
; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v8.h, v52.h
; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.l, v66.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v24.h, v66.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v32.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_hi16
; GFX11-TRUE16-NEXT:  .LBB92_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB92_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v78.l, 3
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v48.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v39.h, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v75.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v36.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v74.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v43.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v71.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v71.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v39.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v70.h, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.l, 8, v46.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v38.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v35.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.l, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v69.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v72.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v69.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v70.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v61.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v35.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v36.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v33.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v59.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v68.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v68.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v34.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v183.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v67.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.h, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v33.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v32.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v67.h, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v180.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v44.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v66.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v151.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v66.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v176.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v167.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v179.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v161.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v150.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v134.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v163.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v145.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v131.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v118.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v149.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v129.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v148.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v144.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v115.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v116.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v103.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v119.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v86.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v112.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v99.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v85.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v102.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v96.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v84.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v82.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v98.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v83.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v81.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v32.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v80.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v3.l, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v51.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v50.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v65.h, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v79.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v51.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v64.l, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v64.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v76.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v50.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v65.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v49.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v77.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v55.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v48.h, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v73.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v49.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v54.h, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v38.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v55.l, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v54.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v63.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v37.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v53.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v56.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v37.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v34.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v53.l, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v58.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v52.l, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v52.h, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v42.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v60.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v62.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v41.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v47.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v57.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v40.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v182.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v166.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v45.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.h, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v178.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v164.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v181.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v133.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v162.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v177.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v165.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v147.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v132.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v135.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v128.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v3.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v3.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v146.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v160.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v113.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v130.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v32.h, v117.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.l, 8, v114.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.l, 8, v101.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.l, 8, v97.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v87.l
; GFX11-TRUE16-NEXT:    v_and_b16 v32.h, 0xff, v32.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.l, 8, v100.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.l, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v32.l, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v33.l, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v33.h, v2.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.h, v34.l, v32.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v32.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v32.h
; GFX11-TRUE16-NEXT:  .LBB92_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:392
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:488
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64f16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, v24 :: v_dual_mov_b32 v51, v19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, v23 :: v_dual_mov_b32 v71, v20
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, v22 :: v_dual_mov_b32 v67, v17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v68, v21 :: v_dual_mov_b32 v65, v18
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, v16 :: v_dual_mov_b32 v50, v15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v14 :: v_dual_mov_b32 v53, v9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v13 :: v_dual_mov_b32 v49, v10
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, v12 :: v_dual_mov_b32 v33, v7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v55, v8
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v39, v6 :: v_dual_mov_b32 v36, v5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, v4 :: v_dual_mov_b32 v37, v2
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v38, v0
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v0, off, s32 offset:384
; GFX11-FAKE16-NEXT:    scratch_load_u16 v105, off, s32 offset:380
; GFX11-FAKE16-NEXT:    scratch_load_u16 v93, off, s32 offset:376
; GFX11-FAKE16-NEXT:    scratch_load_u16 v104, off, s32 offset:372
; GFX11-FAKE16-NEXT:    scratch_load_u16 v1, off, s32 offset:368
; GFX11-FAKE16-NEXT:    scratch_load_u16 v94, off, s32 offset:364
; GFX11-FAKE16-NEXT:    scratch_load_u16 v2, off, s32 offset:360
; GFX11-FAKE16-NEXT:    scratch_load_u16 v106, off, s32 offset:356
; GFX11-FAKE16-NEXT:    scratch_load_u16 v3, off, s32 offset:352
; GFX11-FAKE16-NEXT:    scratch_load_u16 v91, off, s32 offset:348
; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:344
; GFX11-FAKE16-NEXT:    scratch_load_u16 v89, off, s32 offset:340
; GFX11-FAKE16-NEXT:    scratch_load_u16 v4, off, s32 offset:336
; GFX11-FAKE16-NEXT:    scratch_load_u16 v88, off, s32 offset:332
; GFX11-FAKE16-NEXT:    scratch_load_u16 v5, off, s32 offset:328
; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_load_u16 v6, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v62, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v7, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v73, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v8, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v79, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v9, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v63, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v45, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v61, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v10, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v11, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v72, off, s32 offset:260
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v57, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:188
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:132
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v12, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_b32 v13, off, s32 offset:388
; GFX11-FAKE16-NEXT:    scratch_load_u16 v14, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v15, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v16, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v17, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v18, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v19, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v20, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v21, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v22, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v23, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v24, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:12
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v83, v30 :: v_dual_mov_b32 v86, v28
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, v29 :: v_dual_mov_b32 v80, v27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, v26 :: v_dual_mov_b32 v81, v25
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(62)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v127, 8, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v126, 8, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v125, 8, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v123, 8, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v122, 8, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v124, 8, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v121, 8, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v120, 8, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v111, 8, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v110, 8, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v108, 8, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v109, 8, v11
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(33)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v107, 8, v12
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(29)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(28)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v182, 8, v14
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(27)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v166, 8, v15
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(26)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v180, 8, v16
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(25)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v46, 8, v17
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(24)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v47, 8, v18
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(23)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v56, 8, v19
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(22)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v75, 8, v20
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(21)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v59, 8, v21
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(20)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v74, 8, v22
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(19)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v95, 8, v23
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(18)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v90, 8, v24
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB92_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v38, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v37, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v39, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v64, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v52, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v69, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v3, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v55, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v49, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v65, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v84, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v82, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v71, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v70, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v3, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v5, v4, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v7, v6, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v11, v10, 0x5040100
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v118
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v113
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v9, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v86, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v83, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v99
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v7, v182
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v134
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v129
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v180
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v112, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v166
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v7, v46
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v47
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v9, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v132
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v11, v14, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v151
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v10, v12, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v13, v15, 0x5040100
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v56
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v145
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v150
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v11, v75
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v167
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v162
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v130, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v59
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v74
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v148, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v11, v95
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v90
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v116
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v12, v13, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v14, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v15, v18, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v17, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v20, v107
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v115, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v103, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v98, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v131, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v117, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v147, v146, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v133, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v164, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v149, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v18, v17, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v20, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v21, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v24, v23, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v177, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v165, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v183, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v179, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v43, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v40, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v60, v57, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v44, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v72
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v21, v20, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v23, v22, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v25, v24, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v58
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v28, v109
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v79
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v73
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v63
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v108
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v61, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v111
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v120
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v110
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v91
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v28, v27, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v88
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v92
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v94
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v105
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v27, v122
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v106
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v30, v121
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v76, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v124
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v31, v123
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v89, v77, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v27, v125
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v126
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v34, v127
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v104, v93, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v28, v30, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v31, v35, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v33, v36, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v34, v37, 0x5040100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr89
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr91
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr106
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr93
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr105
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr95
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr107
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr109
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr108
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr111
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr120
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr121
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr124
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr122
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr125
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr126
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr127
; GFX11-FAKE16-NEXT:  .LBB92_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB92_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v104, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v105, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v93
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v106, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v94, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v92, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v127, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v89, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v125, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v126, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v89, 0x300, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v1, 8, v77
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v91, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v77, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v88, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v1, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v76, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v62
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v123, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v124, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v78, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v62, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v122, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v5, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v79, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v73, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v73, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v121, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v61, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v72, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v61, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v63, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v120, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v111, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v45
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v45, 0x300, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v110, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v58, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v60, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v42
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v58, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v109, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v57
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v44, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v43, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v108, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v41
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v40, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v40, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v183, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v181, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v179, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v178
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v176
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v177, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v176, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v165, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v163
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v161
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v164, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v161, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v160
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v149, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v147, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v147, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v144
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v133, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v131, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v119
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v128
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v117, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v117, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v103, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v102, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v98, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v97
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v115, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v96, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v116, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(16)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v167, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v97, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v107, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v162, 3
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(13)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v148, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v98, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v95, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v150, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v135
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(11)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v151, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v90, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v145, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v100, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v74, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v75, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(8)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v130, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v103, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v132, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v59, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(6)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v134, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v129, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v114, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v56, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v112, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v112, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v46, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v47, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v113, 3
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v118, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v101, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v99, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v86, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v180, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v182, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v83, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v83, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v166, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v87
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v84, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v84, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v82, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v81
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v80
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v80, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v71, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v70, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v68
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v4, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v69, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v68, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v5, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v66, 8, v67
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v6, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v65, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v64, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v54
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v52, v52, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v67, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v66, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v51, 8, v51
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v50, 8, v50
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v55, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v51, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v49, v49, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v51, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v50, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v48, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v39, v39, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v38, v38, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v37, v37, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v53, 8, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xff, v49
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v36, 8, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v35, 8, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v33, 8, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v53, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v36, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v35, v38
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v34, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v48
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v37, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v34, v35, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v33, v3, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v32, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v36, v51, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v37, v4, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v6, v5, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v68, v67, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v80, v7, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v84, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v83, v9, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v101, v10, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v112, v11, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v114, v12, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v103, v13, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v100, v14, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v98, v15, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v97, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v96, v17, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v102, v18, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v117, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v147, v20, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v161, v21, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v176, v22, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v181, v23, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v40, v24, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v58, v25, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v45, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v61, v27, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v73, v28, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v62, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v77, v30, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v89, v31, 0x5040100
; GFX11-FAKE16-NEXT:  .LBB92_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define inreg <64 x half> @bitcast_v128i8_to_v64f16_scalar(<128 x i8> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v128i8_to_v64f16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    ; implicit-def: $vgpr43 : SGPR spill to VGPR lane
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_writelane_b32 v40, s30, 0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v43, s29, 0
; SI-NEXT:    v_writelane_b32 v43, s28, 1
; SI-NEXT:    v_writelane_b32 v43, s27, 2
; SI-NEXT:    v_writelane_b32 v43, s26, 3
; SI-NEXT:    v_writelane_b32 v43, s25, 4
; SI-NEXT:    v_writelane_b32 v40, s31, 1
; SI-NEXT:    v_writelane_b32 v43, s24, 5
; SI-NEXT:    v_writelane_b32 v40, s34, 2
; SI-NEXT:    v_writelane_b32 v43, s23, 6
; SI-NEXT:    v_writelane_b32 v40, s35, 3
; SI-NEXT:    v_writelane_b32 v43, s22, 7
; SI-NEXT:    v_writelane_b32 v40, s36, 4
; SI-NEXT:    v_writelane_b32 v43, s21, 8
; SI-NEXT:    v_writelane_b32 v40, s37, 5
; SI-NEXT:    v_writelane_b32 v43, s20, 9
; SI-NEXT:    v_writelane_b32 v40, s38, 6
; SI-NEXT:    v_writelane_b32 v43, s19, 10
; SI-NEXT:    v_writelane_b32 v40, s39, 7
; SI-NEXT:    v_writelane_b32 v43, s18, 11
; SI-NEXT:    v_writelane_b32 v40, s48, 8
; SI-NEXT:    v_writelane_b32 v43, s17, 12
; SI-NEXT:    v_writelane_b32 v40, s49, 9
; SI-NEXT:    v_writelane_b32 v43, s16, 13
; SI-NEXT:    v_readfirstlane_b32 s4, v30
; SI-NEXT:    v_writelane_b32 v40, s50, 10
; SI-NEXT:    v_writelane_b32 v43, s4, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v29
; SI-NEXT:    v_writelane_b32 v40, s51, 11
; SI-NEXT:    v_writelane_b32 v43, s4, 15
; SI-NEXT:    v_readfirstlane_b32 s4, v28
; SI-NEXT:    v_writelane_b32 v40, s52, 12
; SI-NEXT:    v_writelane_b32 v43, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v27
; SI-NEXT:    v_writelane_b32 v40, s53, 13
; SI-NEXT:    v_writelane_b32 v43, s4, 17
; SI-NEXT:    v_readfirstlane_b32 s4, v26
; SI-NEXT:    v_writelane_b32 v40, s54, 14
; SI-NEXT:    v_writelane_b32 v43, s4, 18
; SI-NEXT:    v_readfirstlane_b32 s4, v25
; SI-NEXT:    v_writelane_b32 v40, s55, 15
; SI-NEXT:    v_writelane_b32 v43, s4, 19
; SI-NEXT:    v_readfirstlane_b32 s4, v24
; SI-NEXT:    v_writelane_b32 v40, s64, 16
; SI-NEXT:    v_writelane_b32 v43, s4, 20
; SI-NEXT:    v_readfirstlane_b32 s4, v23
; SI-NEXT:    v_writelane_b32 v40, s65, 17
; SI-NEXT:    v_writelane_b32 v43, s4, 21
; SI-NEXT:    v_readfirstlane_b32 s4, v22
; SI-NEXT:    v_writelane_b32 v40, s66, 18
; SI-NEXT:    v_writelane_b32 v43, s4, 22
; SI-NEXT:    v_readfirstlane_b32 s4, v21
; SI-NEXT:    v_writelane_b32 v40, s67, 19
; SI-NEXT:    v_writelane_b32 v43, s4, 23
; SI-NEXT:    v_readfirstlane_b32 s4, v20
; SI-NEXT:    v_writelane_b32 v40, s68, 20
; SI-NEXT:    v_writelane_b32 v43, s4, 24
; SI-NEXT:    v_readfirstlane_b32 s4, v19
; SI-NEXT:    v_writelane_b32 v40, s69, 21
; SI-NEXT:    v_writelane_b32 v43, s4, 25
; SI-NEXT:    v_readfirstlane_b32 s4, v17
; SI-NEXT:    v_writelane_b32 v40, s70, 22
; SI-NEXT:    v_writelane_b32 v43, s4, 26
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_writelane_b32 v40, s71, 23
; SI-NEXT:    v_readfirstlane_b32 s71, v18
; SI-NEXT:    v_writelane_b32 v43, s4, 27
; SI-NEXT:    v_readfirstlane_b32 s4, v15
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:228
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:220
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:316
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:308
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:300
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:292
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:268
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:260
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:252
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:244
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:332
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:236
; SI-NEXT:    v_writelane_b32 v43, s4, 28
; SI-NEXT:    v_readfirstlane_b32 s4, v14
; SI-NEXT:    v_writelane_b32 v43, s4, 29
; SI-NEXT:    v_readfirstlane_b32 s70, v13
; SI-NEXT:    v_readfirstlane_b32 s4, v12
; SI-NEXT:    v_readfirstlane_b32 s25, v11
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204
; SI-NEXT:    v_writelane_b32 v43, s4, 30
; SI-NEXT:    v_readfirstlane_b32 s4, v9
; SI-NEXT:    v_writelane_b32 v43, s4, 31
; SI-NEXT:    v_readfirstlane_b32 s4, v8
; SI-NEXT:    v_writelane_b32 v43, s4, 32
; SI-NEXT:    v_readfirstlane_b32 s4, v7
; SI-NEXT:    v_writelane_b32 v43, s4, 33
; SI-NEXT:    v_readfirstlane_b32 s4, v6
; SI-NEXT:    v_writelane_b32 v43, s4, 34
; SI-NEXT:    v_readfirstlane_b32 s4, v5
; SI-NEXT:    v_writelane_b32 v43, s4, 35
; SI-NEXT:    v_readfirstlane_b32 s4, v4
; SI-NEXT:    v_writelane_b32 v43, s4, 36
; SI-NEXT:    v_readfirstlane_b32 s4, v1
; SI-NEXT:    v_readfirstlane_b32 s37, v2
; SI-NEXT:    v_writelane_b32 v43, s4, 37
; SI-NEXT:    v_readfirstlane_b32 s4, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176
; SI-NEXT:    v_readfirstlane_b32 s24, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:196
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:188
; SI-NEXT:    v_writelane_b32 v40, s80, 24
; SI-NEXT:    v_writelane_b32 v40, s81, 25
; SI-NEXT:    v_writelane_b32 v40, s82, 26
; SI-NEXT:    v_writelane_b32 v40, s83, 27
; SI-NEXT:    v_writelane_b32 v40, s84, 28
; SI-NEXT:    v_writelane_b32 v43, s4, 38
; SI-NEXT:    v_writelane_b32 v40, s85, 29
; SI-NEXT:    v_writelane_b32 v40, s86, 30
; SI-NEXT:    v_writelane_b32 v40, s87, 31
; SI-NEXT:    v_writelane_b32 v40, s96, 32
; SI-NEXT:    v_writelane_b32 v40, s97, 33
; SI-NEXT:    v_writelane_b32 v40, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s28, v3
; SI-NEXT:    ; implicit-def: $vgpr42 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v40, s99, 35
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s56, v19
; SI-NEXT:    v_readfirstlane_b32 s47, v20
; SI-NEXT:    v_readfirstlane_b32 s57, v21
; SI-NEXT:    v_readfirstlane_b32 s58, v22
; SI-NEXT:    v_readfirstlane_b32 s59, v23
; SI-NEXT:    v_readfirstlane_b32 s62, v24
; SI-NEXT:    v_readfirstlane_b32 s89, v25
; SI-NEXT:    v_readfirstlane_b32 s90, v26
; SI-NEXT:    v_readfirstlane_b32 s98, v27
; SI-NEXT:    v_readfirstlane_b32 s65, v28
; SI-NEXT:    v_readfirstlane_b32 s60, v29
; SI-NEXT:    v_readfirstlane_b32 s44, v30
; SI-NEXT:    v_readfirstlane_b32 s46, v31
; SI-NEXT:    v_readfirstlane_b32 s45, v32
; SI-NEXT:    v_readfirstlane_b32 s75, v33
; SI-NEXT:    v_readfirstlane_b32 s79, v34
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:172
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:148
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:132
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:124
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:116
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
; SI-NEXT:    v_readfirstlane_b32 s96, v35
; SI-NEXT:    v_readfirstlane_b32 s84, v36
; SI-NEXT:    v_readfirstlane_b32 s80, v37
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s83, v17
; SI-NEXT:    v_readfirstlane_b32 s21, v16
; SI-NEXT:    v_readfirstlane_b32 s22, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:80
; SI-NEXT:    v_readfirstlane_b32 s92, v38
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s23, v51
; SI-NEXT:    v_readfirstlane_b32 s31, v39
; SI-NEXT:    v_readfirstlane_b32 s97, v14
; SI-NEXT:    v_readfirstlane_b32 s55, v13
; SI-NEXT:    v_readfirstlane_b32 s72, v12
; SI-NEXT:    v_readfirstlane_b32 s73, v11
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:48
; SI-NEXT:    v_readfirstlane_b32 s26, v48
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v49
; SI-NEXT:    v_readfirstlane_b32 s14, v50
; SI-NEXT:    v_writelane_b32 v43, s4, 39
; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
; SI-NEXT:    v_readfirstlane_b32 s82, v1
; SI-NEXT:    v_readfirstlane_b32 s94, v0
; SI-NEXT:    v_readfirstlane_b32 s63, v2
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32
; SI-NEXT:    v_readfirstlane_b32 s86, v10
; SI-NEXT:    v_readfirstlane_b32 s87, v9
; SI-NEXT:    v_readfirstlane_b32 s85, v8
; SI-NEXT:    v_readfirstlane_b32 s69, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:16
; SI-NEXT:    v_readfirstlane_b32 s74, v3
; SI-NEXT:    v_readfirstlane_b32 s61, v4
; SI-NEXT:    v_readfirstlane_b32 s67, v5
; SI-NEXT:    v_readfirstlane_b32 s81, v6
; SI-NEXT:    v_readfirstlane_b32 s29, v19
; SI-NEXT:    v_readfirstlane_b32 s35, v20
; SI-NEXT:    v_readfirstlane_b32 s27, v21
; SI-NEXT:    v_readfirstlane_b32 s17, v22
; SI-NEXT:    v_readfirstlane_b32 s13, v23
; SI-NEXT:    v_readfirstlane_b32 s52, v24
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s51, v25
; SI-NEXT:    v_readfirstlane_b32 s53, v26
; SI-NEXT:    v_readfirstlane_b32 s54, v27
; SI-NEXT:    v_readfirstlane_b32 s66, v28
; SI-NEXT:    v_readfirstlane_b32 s64, v29
; SI-NEXT:    v_readfirstlane_b32 s18, v30
; SI-NEXT:    v_readfirstlane_b32 s12, v15
; SI-NEXT:    v_readfirstlane_b32 s38, v16
; SI-NEXT:    v_readfirstlane_b32 s36, v17
; SI-NEXT:    v_readfirstlane_b32 s39, v18
; SI-NEXT:    v_readfirstlane_b32 s48, v31
; SI-NEXT:    v_readfirstlane_b32 s50, v32
; SI-NEXT:    v_writelane_b32 v42, s50, 0
; SI-NEXT:    v_readfirstlane_b32 s10, v34
; SI-NEXT:    v_writelane_b32 v42, s10, 1
; SI-NEXT:    v_writelane_b32 v42, s51, 2
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_writelane_b32 v42, s52, 3
; SI-NEXT:    v_writelane_b32 v42, s11, 4
; SI-NEXT:    v_writelane_b32 v42, s53, 5
; SI-NEXT:    v_writelane_b32 v42, s54, 6
; SI-NEXT:    v_writelane_b32 v42, s64, 7
; SI-NEXT:    v_writelane_b32 v42, s66, 8
; SI-NEXT:    v_writelane_b32 v42, s67, 9
; SI-NEXT:    v_writelane_b32 v42, s61, 10
; SI-NEXT:    v_writelane_b32 v42, s83, 11
; SI-NEXT:    v_writelane_b32 v42, s86, 12
; SI-NEXT:    v_writelane_b32 v42, s12, 13
; SI-NEXT:    v_writelane_b32 v42, s87, 14
; SI-NEXT:    v_writelane_b32 v42, s81, 15
; SI-NEXT:    v_writelane_b32 v42, s29, 16
; SI-NEXT:    v_writelane_b32 v42, s21, 17
; SI-NEXT:    v_writelane_b32 v42, s85, 18
; SI-NEXT:    v_writelane_b32 v42, s84, 19
; SI-NEXT:    v_writelane_b32 v42, s13, 20
; SI-NEXT:    v_writelane_b32 v42, s65, 21
; SI-NEXT:    v_writelane_b32 v42, s98, 22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_readfirstlane_b32 s91, v1
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_readfirstlane_b32 s16, v2
; SI-NEXT:    v_writelane_b32 v43, s16, 40
; SI-NEXT:    v_writelane_b32 v42, s14, 23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_readfirstlane_b32 s15, v51
; SI-NEXT:    v_writelane_b32 v43, s91, 41
; SI-NEXT:    v_writelane_b32 v42, s55, 24
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_readfirstlane_b32 s19, v9
; SI-NEXT:    v_writelane_b32 v43, s15, 42
; SI-NEXT:    v_writelane_b32 v42, s82, 25
; SI-NEXT:    v_readfirstlane_b32 s20, v8
; SI-NEXT:    v_writelane_b32 v43, s19, 43
; SI-NEXT:    v_writelane_b32 v42, s23, 26
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_readfirstlane_b32 s40, v10
; SI-NEXT:    v_writelane_b32 v43, s20, 44
; SI-NEXT:    v_writelane_b32 v42, s94, 27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_readfirstlane_b32 s41, v39
; SI-NEXT:    v_writelane_b32 v43, s40, 45
; SI-NEXT:    v_writelane_b32 v42, s60, 28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_readfirstlane_b32 s42, v49
; SI-NEXT:    v_writelane_b32 v43, s41, 46
; SI-NEXT:    v_writelane_b32 v42, s44, 29
; SI-NEXT:    v_readfirstlane_b32 s43, v48
; SI-NEXT:    v_writelane_b32 v43, s42, 47
; SI-NEXT:    v_writelane_b32 v42, s72, 30
; SI-NEXT:    v_readfirstlane_b32 s76, v13
; SI-NEXT:    v_writelane_b32 v43, s43, 48
; SI-NEXT:    v_writelane_b32 v42, s63, 31
; SI-NEXT:    v_readfirstlane_b32 s77, v12
; SI-NEXT:    v_writelane_b32 v43, s76, 49
; SI-NEXT:    v_writelane_b32 v42, s73, 32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s6, v50
; SI-NEXT:    v_writelane_b32 v43, s77, 50
; SI-NEXT:    v_writelane_b32 v42, s45, 33
; SI-NEXT:    v_readfirstlane_b32 s7, v0
; SI-NEXT:    v_writelane_b32 v43, s6, 51
; SI-NEXT:    v_writelane_b32 v42, s46, 34
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v14
; SI-NEXT:    v_writelane_b32 v43, s7, 52
; SI-NEXT:    v_writelane_b32 v42, s74, 35
; SI-NEXT:    v_readfirstlane_b32 vcc_hi, v35
; SI-NEXT:    v_writelane_b32 v43, vcc_lo, 53
; SI-NEXT:    v_writelane_b32 v42, s75, 36
; SI-NEXT:    v_readfirstlane_b32 s30, v37
; SI-NEXT:    v_writelane_b32 v43, vcc_hi, 54
; SI-NEXT:    v_writelane_b32 v42, s47, 37
; SI-NEXT:    v_readfirstlane_b32 s34, v36
; SI-NEXT:    v_writelane_b32 v43, s30, 55
; SI-NEXT:    v_writelane_b32 v42, s79, 38
; SI-NEXT:    v_writelane_b32 v43, s34, 56
; SI-NEXT:    v_writelane_b32 v42, s57, 39
; SI-NEXT:    v_writelane_b32 v43, s36, 57
; SI-NEXT:    v_writelane_b32 v42, s58, 40
; SI-NEXT:    v_readfirstlane_b32 s8, v38
; SI-NEXT:    v_writelane_b32 v43, s38, 58
; SI-NEXT:    v_writelane_b32 v42, s56, 41
; SI-NEXT:    v_readfirstlane_b32 s9, v7
; SI-NEXT:    v_writelane_b32 v43, s8, 59
; SI-NEXT:    v_writelane_b32 v42, s89, 42
; SI-NEXT:    v_writelane_b32 v43, s9, 60
; SI-NEXT:    v_writelane_b32 v42, s90, 43
; SI-NEXT:    v_writelane_b32 v43, s39, 61
; SI-NEXT:    v_writelane_b32 v42, s62, 44
; SI-NEXT:    v_readfirstlane_b32 s49, v33
; SI-NEXT:    v_writelane_b32 v43, s48, 62
; SI-NEXT:    v_writelane_b32 v42, s59, 45
; SI-NEXT:    v_writelane_b32 v43, s49, 63
; SI-NEXT:    v_writelane_b32 v42, s96, 46
; SI-NEXT:    s_cbranch_scc0 .LBB93_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_readlane_b32 s4, v43, 13
; SI-NEXT:    v_readlane_b32 s5, v43, 12
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 47
; SI-NEXT:    v_readlane_b32 s4, v43, 5
; SI-NEXT:    v_readlane_b32 s5, v43, 4
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 49
; SI-NEXT:    s_and_b32 s4, s37, 0xff
; SI-NEXT:    s_lshl_b32 s5, s28, 8
; SI-NEXT:    v_writelane_b32 v42, s97, 50
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 51
; SI-NEXT:    s_and_b32 s4, s24, 0xff
; SI-NEXT:    s_lshl_b32 s5, s25, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_readlane_b32 s5, v43, 25
; SI-NEXT:    v_writelane_b32 v42, s4, 52
; SI-NEXT:    s_and_b32 s4, s71, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 53
; SI-NEXT:    v_readlane_b32 s4, v43, 18
; SI-NEXT:    v_readlane_b32 s5, v43, 17
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 54
; SI-NEXT:    s_and_b32 s4, s7, 0xff
; SI-NEXT:    s_lshl_b32 s5, s6, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 55
; SI-NEXT:    s_and_b32 s4, s9, 0xff
; SI-NEXT:    s_lshl_b32 s5, s8, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 56
; SI-NEXT:    s_and_b32 s4, s11, 0xff
; SI-NEXT:    s_lshl_b32 s5, s10, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 57
; SI-NEXT:    s_and_b32 s4, s12, 0xff
; SI-NEXT:    s_lshl_b32 s5, s18, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 58
; SI-NEXT:    s_and_b32 s4, s13, 0xff
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_or_b32 s97, s4, s5
; SI-NEXT:    s_and_b32 s4, s74, 0xff
; SI-NEXT:    s_lshl_b32 s5, s63, 8
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s73, 0xff
; SI-NEXT:    s_lshl_b32 s5, s72, 8
; SI-NEXT:    s_or_b32 s72, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xff
; SI-NEXT:    s_lshl_b32 s5, s14, 8
; SI-NEXT:    s_or_b32 s73, s4, s5
; SI-NEXT:    s_and_b32 s4, s79, 0xff
; SI-NEXT:    s_lshl_b32 s5, s75, 8
; SI-NEXT:    s_or_b32 s74, s4, s5
; SI-NEXT:    s_and_b32 s4, s90, 0xff
; SI-NEXT:    s_lshl_b32 s5, s89, 8
; SI-NEXT:    s_or_b32 s75, s4, s5
; SI-NEXT:    v_readlane_b32 s4, v43, 9
; SI-NEXT:    v_readlane_b32 s5, v43, 8
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s5, s4, s5
; SI-NEXT:    v_readlane_b32 s4, v43, 7
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    v_readlane_b32 s6, v43, 6
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_or_b32 s79, s6, s4
; SI-NEXT:    v_readlane_b32 s4, v43, 11
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    v_readlane_b32 s6, v43, 10
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_or_b32 s4, s6, s4
; SI-NEXT:    v_readlane_b32 s6, v43, 1
; SI-NEXT:    v_readlane_b32 s7, v43, 0
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_or_b32 s7, s6, s7
; SI-NEXT:    v_readlane_b32 s6, v43, 38
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s8, v43, 37
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_or_b32 s89, s8, s6
; SI-NEXT:    v_readlane_b32 s6, v43, 3
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s8, v43, 2
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_readlane_b32 s8, v43, 34
; SI-NEXT:    v_readlane_b32 s9, v43, 33
; SI-NEXT:    v_writelane_b32 v42, s26, 59
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s9, s9, 8
; SI-NEXT:    v_writelane_b32 v42, s31, 60
; SI-NEXT:    s_or_b32 s9, s8, s9
; SI-NEXT:    v_readlane_b32 s8, v43, 32
; SI-NEXT:    v_writelane_b32 v42, s17, 61
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v43, 31
; SI-NEXT:    v_writelane_b32 v42, s35, 62
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    v_writelane_b32 v42, s27, 63
; SI-NEXT:    s_or_b32 s27, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v43, 36
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v43, 35
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v43, 29
; SI-NEXT:    v_readlane_b32 s11, v43, 28
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    s_lshl_b32 s11, s11, 8
; SI-NEXT:    s_or_b32 s11, s10, s11
; SI-NEXT:    v_readlane_b32 s10, v43, 27
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s12, v43, 26
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s12, s12, 24
; SI-NEXT:    s_or_b32 s93, s12, s10
; SI-NEXT:    v_readlane_b32 s10, v43, 30
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s12, s70, 24
; SI-NEXT:    s_or_b32 s10, s12, s10
; SI-NEXT:    v_readlane_b32 s12, v43, 22
; SI-NEXT:    v_readlane_b32 s13, v43, 21
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    s_lshl_b32 s13, s13, 8
; SI-NEXT:    s_or_b32 s13, s12, s13
; SI-NEXT:    v_readlane_b32 s12, v43, 20
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s14, v43, 19
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_or_b32 s26, s14, s12
; SI-NEXT:    v_readlane_b32 s12, v43, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s14, v43, 23
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_or_b32 s12, s14, s12
; SI-NEXT:    v_readlane_b32 s14, v43, 14
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s15, s15, 8
; SI-NEXT:    s_or_b32 s15, s14, s15
; SI-NEXT:    s_and_b32 s14, s16, 0xff
; SI-NEXT:    ; implicit-def: $vgpr41 : SGPR spill to VGPR lane
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    v_writelane_b32 v41, s18, 0
; SI-NEXT:    s_lshl_b32 s18, s91, 24
; SI-NEXT:    s_or_b32 s68, s18, s14
; SI-NEXT:    v_readlane_b32 s14, v43, 16
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s16, v43, 15
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_lshl_b32 s18, s16, 24
; SI-NEXT:    s_or_b32 s14, s18, s14
; SI-NEXT:    s_and_b32 s18, s41, 0xff
; SI-NEXT:    s_lshl_b32 s40, s40, 8
; SI-NEXT:    s_or_b32 s41, s18, s40
; SI-NEXT:    s_and_b32 s18, s19, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_lshl_b32 s40, s20, 24
; SI-NEXT:    s_mov_b32 s95, s24
; SI-NEXT:    s_or_b32 s24, s40, s18
; SI-NEXT:    s_and_b32 s40, s42, 0xff
; SI-NEXT:    s_lshl_b32 s40, s40, 16
; SI-NEXT:    s_lshl_b32 s42, s43, 24
; SI-NEXT:    s_or_b32 s40, s42, s40
; SI-NEXT:    s_and_b32 s42, vcc_hi, 0xff
; SI-NEXT:    s_lshl_b32 s43, vcc_lo, 8
; SI-NEXT:    s_or_b32 s43, s42, s43
; SI-NEXT:    s_and_b32 s42, s76, 0xff
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s76, s77, 24
; SI-NEXT:    s_mov_b32 s35, s69
; SI-NEXT:    s_or_b32 s69, s76, s42
; SI-NEXT:    s_and_b32 s42, s30, 0xff
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s76, s34, 24
; SI-NEXT:    s_or_b32 s42, s76, s42
; SI-NEXT:    s_and_b32 s76, s48, 0xff
; SI-NEXT:    s_lshl_b32 s77, s39, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s36, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s38, 24
; SI-NEXT:    v_writelane_b32 v41, s70, 1
; SI-NEXT:    s_or_b32 s70, s78, s77
; SI-NEXT:    s_and_b32 s77, s49, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s50, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 vcc_lo, s78, s77
; SI-NEXT:    s_or_b32 vcc_hi, s76, s70
; SI-NEXT:    s_and_b32 s76, s54, 0xff
; SI-NEXT:    s_lshl_b32 s77, s53, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s51, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s52, 24
; SI-NEXT:    v_writelane_b32 v41, s71, 2
; SI-NEXT:    s_or_b32 s71, s78, s77
; SI-NEXT:    s_and_b32 s77, s64, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s66, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s38, s78, s77
; SI-NEXT:    s_or_b32 s39, s76, s71
; SI-NEXT:    s_and_b32 s76, s29, 0xff
; SI-NEXT:    s_lshl_b32 s77, s81, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s67, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s61, 24
; SI-NEXT:    v_readlane_b32 s16, v42, 63
; SI-NEXT:    s_mov_b32 s99, s28
; SI-NEXT:    s_mov_b32 s28, s80
; SI-NEXT:    s_or_b32 s80, s78, s77
; SI-NEXT:    s_and_b32 s77, s16, 0xff
; SI-NEXT:    v_readlane_b32 s16, v42, 62
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s48, s78, s77
; SI-NEXT:    s_or_b32 s49, s76, s80
; SI-NEXT:    s_and_b32 s76, s35, 0xff
; SI-NEXT:    s_lshl_b32 s77, s85, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s87, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s86, 24
; SI-NEXT:    s_or_b32 s81, s78, s77
; SI-NEXT:    s_and_b32 s77, s94, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s82, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s50, s78, s77
; SI-NEXT:    s_or_b32 s51, s76, s81
; SI-NEXT:    s_and_b32 s76, s22, 0xff
; SI-NEXT:    s_lshl_b32 s77, s21, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s83, 0xff
; SI-NEXT:    v_readlane_b32 s16, v43, 39
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_or_b32 s82, s78, s77
; SI-NEXT:    s_and_b32 s77, s55, 0xff
; SI-NEXT:    v_readlane_b32 s16, v42, 50
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s52, s78, s77
; SI-NEXT:    s_or_b32 s53, s76, s82
; SI-NEXT:    s_and_b32 s76, s92, 0xff
; SI-NEXT:    s_lshl_b32 s77, s28, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s84, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s96, 24
; SI-NEXT:    v_readlane_b32 s16, v42, 59
; SI-NEXT:    s_or_b32 s20, s78, s77
; SI-NEXT:    s_and_b32 s77, s16, 0xff
; SI-NEXT:    v_readlane_b32 s16, v42, 60
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s54, s78, s77
; SI-NEXT:    s_or_b32 s55, s76, s20
; SI-NEXT:    s_and_b32 s76, s44, 0xff
; SI-NEXT:    s_lshl_b32 s77, s60, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s65, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s98, 24
; SI-NEXT:    v_readlane_b32 s16, v42, 47
; SI-NEXT:    s_or_b32 s21, s78, s77
; SI-NEXT:    s_and_b32 s77, s45, 0xff
; SI-NEXT:    s_and_b32 s44, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 49
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s46, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_and_b32 s45, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 51
; SI-NEXT:    s_or_b32 s64, s78, s77
; SI-NEXT:    s_or_b32 s65, s76, s21
; SI-NEXT:    s_and_b32 s76, s58, 0xff
; SI-NEXT:    s_lshl_b32 s77, s57, 8
; SI-NEXT:    s_and_b32 s46, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 52
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s47, 0xff
; SI-NEXT:    s_and_b32 s47, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 53
; SI-NEXT:    s_lshl_b32 s78, s56, 24
; SI-NEXT:    s_and_b32 s56, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 54
; SI-NEXT:    s_and_b32 s57, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 55
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_and_b32 s30, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 56
; SI-NEXT:    s_or_b32 s7, s7, s89
; SI-NEXT:    s_or_b32 s96, s78, s77
; SI-NEXT:    s_and_b32 s77, s62, 0xff
; SI-NEXT:    s_and_b32 s34, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 57
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s59, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_and_b32 s36, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 58
; SI-NEXT:    s_and_b32 s86, s97, 0xffff
; SI-NEXT:    s_and_b32 s87, s72, 0xffff
; SI-NEXT:    s_and_b32 s97, s73, 0xffff
; SI-NEXT:    s_or_b32 s72, s45, s6
; SI-NEXT:    s_mov_b32 s73, s7
; SI-NEXT:    s_lshr_b64 s[18:19], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[6:7], vcc, 16
; SI-NEXT:    s_or_b32 s5, s5, s79
; SI-NEXT:    s_or_b32 s9, s9, s27
; SI-NEXT:    s_or_b32 s11, s11, s93
; SI-NEXT:    s_or_b32 s13, s13, s26
; SI-NEXT:    s_or_b32 s15, s15, s68
; SI-NEXT:    s_or_b32 s41, s41, s24
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_mov_b32 s29, s99
; SI-NEXT:    s_or_b32 s66, s78, s77
; SI-NEXT:    s_or_b32 s67, s76, s96
; SI-NEXT:    s_and_b32 s83, s16, 0xffff
; SI-NEXT:    s_and_b32 s98, s63, 0xffff
; SI-NEXT:    s_and_b32 s99, s74, 0xffff
; SI-NEXT:    s_and_b32 s78, s75, 0xffff
; SI-NEXT:    v_writelane_b32 v42, s6, 47
; SI-NEXT:    s_mov_b32 s88, s22
; SI-NEXT:    s_or_b32 s43, s43, s69
; SI-NEXT:    s_mov_b32 s31, s95
; SI-NEXT:    s_or_b32 s74, s44, s4
; SI-NEXT:    s_mov_b32 s75, s5
; SI-NEXT:    s_lshr_b64 s[22:23], s[4:5], 16
; SI-NEXT:    s_or_b32 s62, s46, s8
; SI-NEXT:    s_lshr_b64 s[94:95], s[8:9], 16
; SI-NEXT:    s_or_b32 s60, s47, s10
; SI-NEXT:    s_mov_b32 s61, s11
; SI-NEXT:    s_lshr_b64 s[90:91], s[10:11], 16
; SI-NEXT:    s_or_b32 s58, s56, s12
; SI-NEXT:    s_mov_b32 s59, s13
; SI-NEXT:    s_lshr_b64 s[4:5], s[12:13], 16
; SI-NEXT:    s_or_b32 s56, s57, s14
; SI-NEXT:    s_mov_b32 s57, s15
; SI-NEXT:    s_lshr_b64 s[76:77], s[14:15], 16
; SI-NEXT:    s_or_b32 s46, s30, s40
; SI-NEXT:    s_mov_b32 s47, s41
; SI-NEXT:    s_lshr_b64 s[16:17], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v42, s7, 48
; SI-NEXT:    s_or_b32 s40, s83, s38
; SI-NEXT:    s_mov_b32 s41, s39
; SI-NEXT:    s_lshr_b64 s[38:39], s[38:39], 16
; SI-NEXT:    s_or_b32 s14, s86, s48
; SI-NEXT:    s_mov_b32 s15, s49
; SI-NEXT:    s_lshr_b64 s[48:49], s[48:49], 16
; SI-NEXT:    s_or_b32 s12, s98, s50
; SI-NEXT:    s_mov_b32 s13, s51
; SI-NEXT:    s_lshr_b64 s[50:51], s[50:51], 16
; SI-NEXT:    s_or_b32 s10, s87, s52
; SI-NEXT:    s_mov_b32 s11, s53
; SI-NEXT:    s_lshr_b64 s[52:53], s[52:53], 16
; SI-NEXT:    s_or_b32 s6, s99, s64
; SI-NEXT:    s_or_b32 s98, s78, s66
; SI-NEXT:    s_mov_b32 s99, s67
; SI-NEXT:    s_lshr_b64 s[66:67], s[66:67], 16
; SI-NEXT:    s_mov_b32 s63, s9
; SI-NEXT:    s_or_b32 s44, s34, s42
; SI-NEXT:    s_mov_b32 s45, s43
; SI-NEXT:    s_lshr_b64 s[84:85], s[42:43], 16
; SI-NEXT:    s_or_b32 s42, s36, vcc_lo
; SI-NEXT:    s_mov_b32 s43, vcc_hi
; SI-NEXT:    s_or_b32 s8, s97, s54
; SI-NEXT:    s_mov_b32 s9, s55
; SI-NEXT:    s_lshr_b64 s[54:55], s[54:55], 16
; SI-NEXT:    v_readlane_b32 s97, v42, 50
; SI-NEXT:    s_mov_b32 s7, s65
; SI-NEXT:    s_lshr_b64 s[64:65], s[64:65], 16
; SI-NEXT:    s_lshr_b32 s19, s79, 16
; SI-NEXT:    s_lshr_b32 s67, s89, 16
; SI-NEXT:    s_lshr_b32 s5, s27, 16
; SI-NEXT:    s_lshr_b32 s79, s93, 16
; SI-NEXT:    s_lshr_b32 s89, s26, 16
; SI-NEXT:    s_lshr_b32 s91, s68, 16
; SI-NEXT:    s_lshr_b32 s93, s24, 16
; SI-NEXT:    s_mov_b32 s86, s18
; SI-NEXT:    s_lshr_b32 s95, s69, 16
; SI-NEXT:    s_lshr_b32 s30, s70, 16
; SI-NEXT:    v_readlane_b32 s70, v41, 1
; SI-NEXT:    s_lshr_b32 s34, s71, 16
; SI-NEXT:    s_mov_b32 s24, s31
; SI-NEXT:    v_readlane_b32 s71, v41, 2
; SI-NEXT:    s_lshr_b32 s36, s80, 16
; SI-NEXT:    s_lshr_b32 s39, s81, 16
; SI-NEXT:    s_lshr_b32 s49, s82, 16
; SI-NEXT:    s_mov_b32 s82, s4
; SI-NEXT:    s_mov_b32 s4, s22
; SI-NEXT:    s_lshr_b32 s51, s20, 16
; SI-NEXT:    v_readlane_b32 s18, v41, 0
; SI-NEXT:    s_lshr_b32 s53, s21, 16
; SI-NEXT:    s_mov_b32 s69, s35
; SI-NEXT:    v_readlane_b32 s27, v42, 63
; SI-NEXT:    v_readlane_b32 s35, v42, 62
; SI-NEXT:    v_readlane_b32 s17, v42, 61
; SI-NEXT:    v_readlane_b32 s31, v42, 60
; SI-NEXT:    v_readlane_b32 s26, v42, 59
; SI-NEXT:    s_mov_b32 s80, s28
; SI-NEXT:    s_mov_b32 s22, s88
; SI-NEXT:    s_lshr_b32 s68, s96, 16
; SI-NEXT:    s_mov_b32 s28, s29
; SI-NEXT:    s_cbranch_execnz .LBB93_3
; SI-NEXT:  .LBB93_2: ; %cmp.true
; SI-NEXT:    v_readlane_b32 s4, v42, 43
; SI-NEXT:    s_add_i32 s19, s4, 3
; SI-NEXT:    v_readlane_b32 s5, v42, 42
; SI-NEXT:    v_readlane_b32 s6, v42, 44
; SI-NEXT:    s_and_b32 s4, s19, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_add_i32 s20, s6, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readlane_b32 s5, v42, 45
; SI-NEXT:    s_and_b32 s6, s20, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 24
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_addk_i32 s4, 0x300
; SI-NEXT:    s_or_b32 s5, s5, s6
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readlane_b32 s5, v42, 40
; SI-NEXT:    s_add_i32 s23, s5, 3
; SI-NEXT:    v_readlane_b32 s6, v42, 39
; SI-NEXT:    v_readlane_b32 s7, v42, 37
; SI-NEXT:    s_and_b32 s5, s23, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_readlane_b32 s6, v42, 41
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_addk_i32 s5, 0x300
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_readlane_b32 s6, v42, 38
; SI-NEXT:    s_add_i32 s68, s6, 3
; SI-NEXT:    v_readlane_b32 s7, v42, 36
; SI-NEXT:    v_readlane_b32 s8, v42, 33
; SI-NEXT:    s_and_b32 s6, s68, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_add_i32 s87, s8, 3
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readlane_b32 s7, v42, 34
; SI-NEXT:    s_and_b32 s8, s87, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 24
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_addk_i32 s6, 0x300
; SI-NEXT:    s_or_b32 s7, s7, s8
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readlane_b32 s7, v42, 29
; SI-NEXT:    s_add_i32 s83, s7, 3
; SI-NEXT:    v_readlane_b32 s8, v42, 28
; SI-NEXT:    v_readlane_b32 s9, v42, 21
; SI-NEXT:    s_and_b32 s7, s83, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_readlane_b32 s8, v42, 22
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_addk_i32 s7, 0x300
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_readlane_b32 s8, v42, 26
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    v_readlane_b32 s9, v42, 23
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s9, s9, 8
; SI-NEXT:    s_add_i32 s10, s26, 3
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    s_lshl_b32 s9, s31, 24
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s9, s9, s10
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_add_i32 s9, s92, 3
; SI-NEXT:    v_readlane_b32 s11, v42, 19
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s10, s80, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s98, s4, 0x3000000
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_readlane_b32 s4, v42, 46
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s10, s4, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s9, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_readlane_b32 s10, v42, 32
; SI-NEXT:    s_add_i32 s86, s10, 3
; SI-NEXT:    v_readlane_b32 s11, v42, 30
; SI-NEXT:    v_readlane_b32 s12, v42, 24
; SI-NEXT:    s_and_b32 s10, s86, 0xff
; SI-NEXT:    s_lshl_b32 s11, s11, 8
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    s_lshl_b32 s11, s97, 24
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_addk_i32 s10, 0x300
; SI-NEXT:    s_or_b32 s11, s11, s12
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    s_add_i32 s11, s22, 3
; SI-NEXT:    v_readlane_b32 s12, v42, 17
; SI-NEXT:    v_readlane_b32 s13, v42, 11
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_readlane_b32 s12, v43, 39
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 24
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_addk_i32 s11, 0x300
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_readlane_b32 s12, v42, 35
; SI-NEXT:    s_add_i32 s97, s12, 3
; SI-NEXT:    v_readlane_b32 s13, v42, 31
; SI-NEXT:    v_readlane_b32 s14, v42, 27
; SI-NEXT:    s_and_b32 s12, s97, 0xff
; SI-NEXT:    s_lshl_b32 s13, s13, 8
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    v_readlane_b32 s13, v42, 25
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s13, s13, 24
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_addk_i32 s12, 0x300
; SI-NEXT:    s_or_b32 s13, s13, s14
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    s_add_i32 s13, s69, 3
; SI-NEXT:    v_readlane_b32 s14, v42, 18
; SI-NEXT:    v_readlane_b32 s15, v42, 14
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 8
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_readlane_b32 s14, v42, 12
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_addk_i32 s13, 0x300
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_readlane_b32 s14, v42, 20
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s15, s17, 8
; SI-NEXT:    s_add_i32 s16, s27, 3
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s15, s35, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_addk_i32 s14, 0x300
; SI-NEXT:    s_or_b32 s15, s15, s16
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    v_readlane_b32 s15, v42, 16
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    v_readlane_b32 s16, v42, 15
; SI-NEXT:    v_readlane_b32 s17, v42, 9
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    v_readlane_b32 s16, v42, 10
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s16, s16, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_addk_i32 s15, 0x300
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    v_readlane_b32 s16, v42, 13
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_lshl_b32 s17, s18, 8
; SI-NEXT:    v_readlane_b32 s18, v42, 7
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 8
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 6
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    v_readlane_b32 s18, v42, 5
; SI-NEXT:    v_readlane_b32 s19, v42, 2
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 8
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_readlane_b32 s18, v42, 3
; SI-NEXT:    s_and_b32 s19, s19, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_add_i32 s40, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v42, 4
; SI-NEXT:    s_add_i32 s41, s17, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v42, 1
; SI-NEXT:    v_readlane_b32 s18, v43, 63
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 0
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 62
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    v_readlane_b32 s18, v43, 61
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 8
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_readlane_b32 s18, v43, 57
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s42, s16, 0x3000000
; SI-NEXT:    s_and_b32 s16, s17, 0xffff
; SI-NEXT:    v_readlane_b32 s17, v43, 58
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s43, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 60
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 59
; SI-NEXT:    v_readlane_b32 s18, v43, 55
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 56
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s44, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 54
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 53
; SI-NEXT:    v_readlane_b32 s18, v43, 49
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 50
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s45, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 52
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 51
; SI-NEXT:    v_readlane_b32 s18, v43, 47
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 48
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s46, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 46
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 45
; SI-NEXT:    v_readlane_b32 s18, v43, 43
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 44
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s47, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 18
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 17
; SI-NEXT:    v_readlane_b32 s18, v43, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 15
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s56, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 14
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 42
; SI-NEXT:    v_readlane_b32 s18, v43, 40
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 41
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s57, s16, 0x3000000
; SI-NEXT:    s_add_i32 s16, s71, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 25
; SI-NEXT:    v_readlane_b32 s18, v43, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 23
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s58, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 22
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 21
; SI-NEXT:    v_readlane_b32 s18, v43, 20
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 19
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s59, s16, 0x3000000
; SI-NEXT:    s_add_i32 s16, s24, 3
; SI-NEXT:    v_readlane_b32 s18, v43, 30
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s25, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s70, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s60, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 29
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 28
; SI-NEXT:    v_readlane_b32 s18, v43, 27
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 26
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s61, s16, 0x3000000
; SI-NEXT:    s_add_i32 s16, s37, 3
; SI-NEXT:    v_readlane_b32 s18, v43, 36
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s28, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 35
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s62, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 34
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 33
; SI-NEXT:    v_readlane_b32 s18, v43, 32
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 31
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s63, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 5
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 4
; SI-NEXT:    v_readlane_b32 s18, v43, 3
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 2
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s72, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 1
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 0
; SI-NEXT:    v_readlane_b32 s18, v43, 38
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 37
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s73, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 13
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 12
; SI-NEXT:    v_readlane_b32 s18, v43, 11
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 10
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s74, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 9
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 8
; SI-NEXT:    v_readlane_b32 s18, v43, 7
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 6
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s99, s5, 0x3000000
; SI-NEXT:    s_add_i32 s10, s10, 0x3000000
; SI-NEXT:    s_add_i32 s11, s11, 0x3000000
; SI-NEXT:    s_add_i32 s12, s12, 0x3000000
; SI-NEXT:    s_add_i32 s13, s13, 0x3000000
; SI-NEXT:    s_add_i32 s14, s14, 0x3000000
; SI-NEXT:    s_add_i32 s15, s15, 0x3000000
; SI-NEXT:    s_add_i32 s75, s16, 0x3000000
; SI-NEXT:    s_lshr_b64 s[18:19], s[42:43], 16
; SI-NEXT:    s_add_i32 s6, s6, 0x3000000
; SI-NEXT:    s_add_i32 s7, s7, 0x3000000
; SI-NEXT:    s_add_i32 s8, s8, 0x3000000
; SI-NEXT:    s_add_i32 s9, s9, 0x3000000
; SI-NEXT:    s_lshr_b64 s[4:5], s[74:75], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[62:63], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[60:61], 16
; SI-NEXT:    v_writelane_b32 v42, s18, 47
; SI-NEXT:    s_lshr_b64 s[38:39], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[48:49], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[50:51], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[52:53], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[66:67], s[98:99], 16
; SI-NEXT:    s_lshr_b64 s[86:87], s[72:73], 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[58:59], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[56:57], 16
; SI-NEXT:    s_lshr_b64 s[16:17], s[46:47], 16
; SI-NEXT:    s_lshr_b64 s[84:85], s[44:45], 16
; SI-NEXT:    v_writelane_b32 v42, s19, 48
; SI-NEXT:    s_lshr_b64 s[54:55], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[64:65], s[6:7], 16
; SI-NEXT:    s_lshr_b32 s19, s75, 16
; SI-NEXT:    s_lshr_b32 s67, s73, 16
; SI-NEXT:    s_lshr_b32 s5, s63, 16
; SI-NEXT:    s_lshr_b32 s79, s61, 16
; SI-NEXT:    s_lshr_b32 s89, s59, 16
; SI-NEXT:    s_lshr_b32 s91, s57, 16
; SI-NEXT:    s_lshr_b32 s93, s47, 16
; SI-NEXT:    s_lshr_b32 s95, s45, 16
; SI-NEXT:    s_lshr_b32 s30, s43, 16
; SI-NEXT:    s_lshr_b32 s34, s41, 16
; SI-NEXT:    s_lshr_b32 s36, s15, 16
; SI-NEXT:    s_lshr_b32 s39, s13, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s51, s9, 16
; SI-NEXT:    s_lshr_b32 s53, s7, 16
; SI-NEXT:    s_lshr_b32 s68, s99, 16
; SI-NEXT:  .LBB93_3: ; %end
; SI-NEXT:    s_and_b32 s17, s74, 0xffff
; SI-NEXT:    s_lshl_b32 s18, s4, 16
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_and_b32 s18, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_and_b32 s19, s72, 0xffff
; SI-NEXT:    s_lshl_b32 s20, s86, 16
; SI-NEXT:    s_or_b32 s19, s19, s20
; SI-NEXT:    s_and_b32 s20, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s21, s67, 16
; SI-NEXT:    s_or_b32 s20, s20, s21
; SI-NEXT:    s_and_b32 s21, s62, 0xffff
; SI-NEXT:    s_lshl_b32 s22, s94, 16
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_and_b32 s22, s63, 0xffff
; SI-NEXT:    s_lshl_b32 s23, s5, 16
; SI-NEXT:    s_or_b32 s22, s22, s23
; SI-NEXT:    s_and_b32 s23, s60, 0xffff
; SI-NEXT:    s_lshl_b32 s24, s90, 16
; SI-NEXT:    s_or_b32 s23, s23, s24
; SI-NEXT:    s_and_b32 s24, s61, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s79, 16
; SI-NEXT:    s_or_b32 s24, s24, s25
; SI-NEXT:    s_and_b32 s25, s58, 0xffff
; SI-NEXT:    s_lshl_b32 s26, s82, 16
; SI-NEXT:    s_or_b32 s25, s25, s26
; SI-NEXT:    s_and_b32 s26, s59, 0xffff
; SI-NEXT:    s_lshl_b32 s27, s89, 16
; SI-NEXT:    s_or_b32 s26, s26, s27
; SI-NEXT:    s_and_b32 s27, s56, 0xffff
; SI-NEXT:    s_lshl_b32 s28, s76, 16
; SI-NEXT:    s_or_b32 s27, s27, s28
; SI-NEXT:    s_and_b32 s28, s57, 0xffff
; SI-NEXT:    s_lshl_b32 s29, s91, 16
; SI-NEXT:    s_or_b32 s28, s28, s29
; SI-NEXT:    s_and_b32 s29, s46, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s16, 16
; SI-NEXT:    s_or_b32 s29, s29, s46
; SI-NEXT:    s_and_b32 s46, s47, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s93, 16
; SI-NEXT:    s_or_b32 s46, s46, s47
; SI-NEXT:    s_and_b32 s44, s44, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s84, 16
; SI-NEXT:    s_or_b32 s44, s44, s47
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s95, 16
; SI-NEXT:    v_readlane_b32 s4, v42, 47
; SI-NEXT:    s_or_b32 s45, s45, s47
; SI-NEXT:    s_and_b32 s42, s42, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s4, 16
; SI-NEXT:    s_or_b32 s42, s42, s47
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s30, 16
; SI-NEXT:    s_or_b32 s43, s43, s47
; SI-NEXT:    s_and_b32 s40, s40, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s38, 16
; SI-NEXT:    s_or_b32 s40, s40, s47
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s34, 16
; SI-NEXT:    s_or_b32 s41, s41, s47
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s48, 16
; SI-NEXT:    s_or_b32 s14, s14, s47
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s36, 16
; SI-NEXT:    s_or_b32 s15, s15, s47
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s50, 16
; SI-NEXT:    s_or_b32 s12, s12, s47
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s39, 16
; SI-NEXT:    s_or_b32 s13, s13, s47
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s52, 16
; SI-NEXT:    s_or_b32 s10, s10, s47
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s49, 16
; SI-NEXT:    s_or_b32 s11, s11, s47
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s54, 16
; SI-NEXT:    s_or_b32 s8, s8, s47
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s51, 16
; SI-NEXT:    s_or_b32 s9, s9, s47
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s64, 16
; SI-NEXT:    v_readlane_b32 s5, v42, 48
; SI-NEXT:    s_or_b32 s6, s6, s47
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s53, 16
; SI-NEXT:    s_or_b32 s7, s7, s47
; SI-NEXT:    s_and_b32 s4, s98, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s66, 16
; SI-NEXT:    s_and_b32 s5, s99, 0xffff
; SI-NEXT:    s_lshl_b32 s16, s68, 16
; SI-NEXT:    s_or_b32 s4, s4, s47
; SI-NEXT:    s_or_b32 s5, s5, s16
; SI-NEXT:    v_mov_b32_e32 v0, s17
; SI-NEXT:    v_mov_b32_e32 v1, s18
; SI-NEXT:    v_mov_b32_e32 v2, s19
; SI-NEXT:    v_mov_b32_e32 v3, s20
; SI-NEXT:    v_mov_b32_e32 v4, s21
; SI-NEXT:    v_mov_b32_e32 v5, s22
; SI-NEXT:    v_mov_b32_e32 v6, s23
; SI-NEXT:    v_mov_b32_e32 v7, s24
; SI-NEXT:    v_mov_b32_e32 v8, s25
; SI-NEXT:    v_mov_b32_e32 v9, s26
; SI-NEXT:    v_mov_b32_e32 v10, s27
; SI-NEXT:    v_mov_b32_e32 v11, s28
; SI-NEXT:    v_mov_b32_e32 v12, s29
; SI-NEXT:    v_mov_b32_e32 v13, s46
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_readlane_b32 s99, v40, 35
; SI-NEXT:    v_readlane_b32 s98, v40, 34
; SI-NEXT:    v_readlane_b32 s97, v40, 33
; SI-NEXT:    v_readlane_b32 s96, v40, 32
; SI-NEXT:    v_readlane_b32 s87, v40, 31
; SI-NEXT:    v_readlane_b32 s86, v40, 30
; SI-NEXT:    v_readlane_b32 s85, v40, 29
; SI-NEXT:    v_readlane_b32 s84, v40, 28
; SI-NEXT:    v_readlane_b32 s83, v40, 27
; SI-NEXT:    v_readlane_b32 s82, v40, 26
; SI-NEXT:    v_readlane_b32 s81, v40, 25
; SI-NEXT:    v_readlane_b32 s80, v40, 24
; SI-NEXT:    v_readlane_b32 s71, v40, 23
; SI-NEXT:    v_readlane_b32 s70, v40, 22
; SI-NEXT:    v_readlane_b32 s69, v40, 21
; SI-NEXT:    v_readlane_b32 s68, v40, 20
; SI-NEXT:    v_readlane_b32 s67, v40, 19
; SI-NEXT:    v_readlane_b32 s66, v40, 18
; SI-NEXT:    v_readlane_b32 s65, v40, 17
; SI-NEXT:    v_readlane_b32 s64, v40, 16
; SI-NEXT:    v_readlane_b32 s55, v40, 15
; SI-NEXT:    v_readlane_b32 s54, v40, 14
; SI-NEXT:    v_readlane_b32 s53, v40, 13
; SI-NEXT:    v_readlane_b32 s52, v40, 12
; SI-NEXT:    v_readlane_b32 s51, v40, 11
; SI-NEXT:    v_readlane_b32 s50, v40, 10
; SI-NEXT:    v_readlane_b32 s49, v40, 9
; SI-NEXT:    v_readlane_b32 s48, v40, 8
; SI-NEXT:    v_readlane_b32 s39, v40, 7
; SI-NEXT:    v_readlane_b32 s38, v40, 6
; SI-NEXT:    v_readlane_b32 s37, v40, 5
; SI-NEXT:    v_readlane_b32 s36, v40, 4
; SI-NEXT:    v_readlane_b32 s35, v40, 3
; SI-NEXT:    v_readlane_b32 s34, v40, 2
; SI-NEXT:    v_readlane_b32 s31, v40, 1
; SI-NEXT:    v_readlane_b32 s30, v40, 0
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB93_4:
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    v_writelane_b32 v42, s6, 47
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr19
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr5
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr16
; SI-NEXT:    ; implicit-def: $sgpr93
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    v_writelane_b32 v42, s7, 48
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    s_branch .LBB93_2
;
; VI-LABEL: bitcast_v128i8_to_v64f16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s63, v12
; VI-NEXT:    v_readfirstlane_b32 s8, v0
; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:280
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:12
; VI-NEXT:    v_readfirstlane_b32 s59, v29
; VI-NEXT:    v_readfirstlane_b32 s47, v27
; VI-NEXT:    v_readfirstlane_b32 s75, v30
; VI-NEXT:    v_readfirstlane_b32 s72, v28
; VI-NEXT:    v_readfirstlane_b32 s60, v26
; VI-NEXT:    v_readfirstlane_b32 s43, v25
; VI-NEXT:    v_readfirstlane_b32 s56, v24
; VI-NEXT:    v_readfirstlane_b32 s15, v23
; VI-NEXT:    v_readfirstlane_b32 s44, v22
; VI-NEXT:    v_readfirstlane_b32 s12, v21
; VI-NEXT:    v_readfirstlane_b32 s40, v20
; VI-NEXT:    v_readfirstlane_b32 s76, v18
; VI-NEXT:    v_readfirstlane_b32 s73, v16
; VI-NEXT:    v_readfirstlane_b32 s9, v19
; VI-NEXT:    v_readfirstlane_b32 s61, v17
; VI-NEXT:    v_readfirstlane_b32 s62, v15
; VI-NEXT:    v_readfirstlane_b32 s74, v14
; VI-NEXT:    v_readfirstlane_b32 s57, v13
; VI-NEXT:    v_readfirstlane_b32 s58, v10
; VI-NEXT:    v_readfirstlane_b32 s41, v9
; VI-NEXT:    v_readfirstlane_b32 s46, v8
; VI-NEXT:    v_readfirstlane_b32 s13, v7
; VI-NEXT:    v_readfirstlane_b32 s42, v6
; VI-NEXT:    v_readfirstlane_b32 s10, v5
; VI-NEXT:    v_readfirstlane_b32 s14, v4
; VI-NEXT:    v_readfirstlane_b32 s7, v3
; VI-NEXT:    v_readfirstlane_b32 s11, v2
; VI-NEXT:    v_readfirstlane_b32 s6, v1
; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:328
; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:316
; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:312
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:308
; VI-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:304
; VI-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:300
; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:292
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:288
; VI-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:272
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:268
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:252
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    s_and_b64 s[4:5], vcc, exec
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:216
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:212
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:208
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:200
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:192
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:188
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:184
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:180
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:176
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:172
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:168
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:164
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:160
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:156
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:152
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:148
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:144
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:140
; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:136
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:132
; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:112
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:64
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:40
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    s_cbranch_scc0 .LBB93_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_mov_b32_e32 v3, s23
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v2, s21
; VI-NEXT:    v_perm_b32 v3, s22, v3, v11
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_perm_b32 v2, s20, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_mov_b32_e32 v4, s6
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    v_perm_b32 v4, s8, v4, v11
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_mov_b32_e32 v5, s10
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_mov_b32_e32 v4, s7
; VI-NEXT:    v_perm_b32 v5, s14, v5, v11
; VI-NEXT:    v_perm_b32 v4, s11, v4, v11
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_mov_b32_e32 v6, s41
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_mov_b32_e32 v5, s13
; VI-NEXT:    v_perm_b32 v6, s46, v6, v11
; VI-NEXT:    v_perm_b32 v5, s42, v5, v11
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_mov_b32_e32 v7, s57
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_mov_b32_e32 v6, s45
; VI-NEXT:    v_perm_b32 v7, s63, v7, v11
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, s61
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    v_mov_b32_e32 v7, s62
; VI-NEXT:    v_perm_b32 v8, s73, v8, v11
; VI-NEXT:    v_perm_b32 v7, s74, v7, v11
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v9, s12
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_mov_b32_e32 v8, s9
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_perm_b32 v8, s76, v8, v11
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_mov_b32_e32 v43, v10
; VI-NEXT:    v_mov_b32_e32 v10, s43
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_perm_b32 v10, s56, v10, v11
; VI-NEXT:    v_perm_b32 v9, s44, v9, v11
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_mov_b32_e32 v35, v12
; VI-NEXT:    v_mov_b32_e32 v12, s59
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_mov_b32_e32 v10, s47
; VI-NEXT:    v_perm_b32 v12, s72, v12, v11
; VI-NEXT:    v_perm_b32 v10, s60, v10, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_or_b32_e32 v10, v10, v12
; VI-NEXT:    v_perm_b32 v12, v15, v14, s4
; VI-NEXT:    v_perm_b32 v11, s75, v13, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v18, v23, v18, s4
; VI-NEXT:    v_perm_b32 v23, v51, v33, s4
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_perm_b32 v12, v56, v12, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v13, v47, v13, s4
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v14, v59, v14, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v15, v32, v15, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v61, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_perm_b32 v14, v17, v60, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_perm_b32 v15, v21, v16, s4
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v17, v28, v24, s4
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v21, v58, v46, s4
; VI-NEXT:    v_perm_b32 v24, v36, v49, s4
; VI-NEXT:    v_mov_b32_e32 v28, v52
; VI-NEXT:    v_mov_b32_e32 v52, v53
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v19, v16, s4
; VI-NEXT:    v_perm_b32 v19, v26, v22, s4
; VI-NEXT:    v_or_b32_e32 v17, v19, v17
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    v_perm_b32 v16, v25, v20, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_or_b32_e32 v16, v18, v16
; VI-NEXT:    v_perm_b32 v18, v45, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v20, v44, v29, s4
; VI-NEXT:    v_perm_b32 v22, v53, v42, s4
; VI-NEXT:    v_mov_b32_e32 v45, v34
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_perm_b32 v19, v30, v19, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_perm_b32 v19, v55, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_perm_b32 v20, v39, v57, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v20, v21, v20
; VI-NEXT:    v_perm_b32 v21, v38, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v21, v22, v21
; VI-NEXT:    v_perm_b32 v22, v54, v37, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v22, v23, v22
; VI-NEXT:    v_perm_b32 v23, v40, v50, s4
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v23, v24, v23
; VI-NEXT:    v_perm_b32 v24, v48, v34, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_mov_b32_e32 v48, v25
; VI-NEXT:    v_perm_b32 v25, v25, v41, s4
; VI-NEXT:    v_or_b32_e32 v24, v25, v24
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v34, v33
; VI-NEXT:    v_mov_b32_e32 v33, v26
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v55, v57
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_mov_b32_e32 v53, v25
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    v_or_b32_e32 v25, v26, v25
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v29, v27, s4
; VI-NEXT:    v_or_b32_e32 v26, v27, v26
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v29, v27, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    v_or_b32_e32 v27, v28, v27
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v28, v28, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v62, v29, s4
; VI-NEXT:    v_or_b32_e32 v28, v29, v28
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, v44
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v43, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v29, v30, v29
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v35, v31, s4
; VI-NEXT:    v_or_b32_e32 v30, v31, v30
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v35, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_mov_b32_e32 v35, v43
; VI-NEXT:    v_perm_b32 v43, v44, v43, s4
; VI-NEXT:    v_or_b32_e32 v31, v43, v31
; VI-NEXT:    v_mov_b32_e32 v43, v39
; VI-NEXT:    v_mov_b32_e32 v39, v34
; VI-NEXT:    v_mov_b32_e32 v34, v52
; VI-NEXT:    v_mov_b32_e32 v52, v45
; VI-NEXT:    s_cbranch_execnz .LBB93_3
; VI-NEXT:  .LBB93_2: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_add_i32 s75, s75, 3
; VI-NEXT:    s_add_i32 s72, s72, 3
; VI-NEXT:    s_add_i32 s60, s60, 3
; VI-NEXT:    s_add_i32 s56, s56, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s76, s76, 3
; VI-NEXT:    s_add_i32 s73, s73, 3
; VI-NEXT:    s_add_i32 s74, s74, 3
; VI-NEXT:    s_add_i32 s63, s63, 3
; VI-NEXT:    s_add_i32 s58, s58, 3
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    v_mov_b32_e32 v18, s17
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v62
; VI-NEXT:    v_perm_b32 v0, v0, v35, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, s21
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v35, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v0, v1, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v33
; VI-NEXT:    v_perm_b32 v25, v0, v53, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x300, v28
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_sdwa v26, v26, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x3000000, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x300, v29
; VI-NEXT:    v_or_b32_sdwa v25, v25, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x3000000, v25
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_perm_b32 v24, v0, v52, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v48
; VI-NEXT:    v_perm_b32 v30, v0, v41, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v40
; VI-NEXT:    v_perm_b32 v23, v0, v50, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v36
; VI-NEXT:    v_perm_b32 v31, v0, v49, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v54
; VI-NEXT:    v_perm_b32 v22, v0, v37, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v51
; VI-NEXT:    v_perm_b32 v33, v0, v39, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v38
; VI-NEXT:    v_perm_b32 v21, v0, v63, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v34
; VI-NEXT:    v_perm_b32 v36, v0, v42, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v43
; VI-NEXT:    v_perm_b32 v20, v0, v55, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v58
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v63, s19
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x300, v30
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x300, v31
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x300, v33
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v23, v23, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v24, v24, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x3000000, v22
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x3000000, v23
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x3000000, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v52, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v53, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v37, vcc, 0x300, v53
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v38, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v34, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v10, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v14, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x300, v15
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_perm_b32 v16, v0, v60, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v61
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x300, v34
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v54, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v59
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v40, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v56
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v42, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v47
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v43, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v44, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v0, 0xc0c0004
; VI-NEXT:    v_perm_b32 v62, s20, v62, v0
; VI-NEXT:    v_perm_b32 v63, s18, v63, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v45, s75, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s59
; VI-NEXT:    v_perm_b32 v55, s72, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_perm_b32 v41, s60, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s43
; VI-NEXT:    v_perm_b32 v9, s56, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s15
; VI-NEXT:    v_perm_b32 v46, s44, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s12
; VI-NEXT:    v_perm_b32 v8, s40, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s9
; VI-NEXT:    v_perm_b32 v47, s76, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s61
; VI-NEXT:    v_perm_b32 v7, s73, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s62
; VI-NEXT:    v_perm_b32 v56, s74, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s57
; VI-NEXT:    v_perm_b32 v6, s63, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s45
; VI-NEXT:    v_perm_b32 v57, s58, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s41
; VI-NEXT:    v_perm_b32 v5, s46, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s13
; VI-NEXT:    v_perm_b32 v58, s42, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s10
; VI-NEXT:    v_perm_b32 v4, s14, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s7
; VI-NEXT:    v_perm_b32 v59, s11, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_perm_b32 v3, s8, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s29
; VI-NEXT:    v_perm_b32 v60, s28, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s27
; VI-NEXT:    v_perm_b32 v2, s26, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s25
; VI-NEXT:    v_perm_b32 v61, s24, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s23
; VI-NEXT:    v_perm_b32 v1, s22, v1, v0
; VI-NEXT:    v_perm_b32 v0, s16, v18, v0
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v63
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
; VI-NEXT:    v_or_b32_sdwa v0, v18, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v62
; VI-NEXT:    v_or_b32_sdwa v1, v1, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v61
; VI-NEXT:    v_or_b32_sdwa v2, v2, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v60
; VI-NEXT:    v_or_b32_sdwa v3, v3, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v59
; VI-NEXT:    v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v58
; VI-NEXT:    v_or_b32_sdwa v5, v5, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v57
; VI-NEXT:    v_or_b32_sdwa v6, v6, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v56
; VI-NEXT:    v_or_b32_sdwa v7, v7, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v47
; VI-NEXT:    v_or_b32_sdwa v8, v8, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v46
; VI-NEXT:    v_or_b32_sdwa v9, v9, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
; VI-NEXT:    v_add_u32_e32 v46, vcc, 0x300, v11
; VI-NEXT:    v_add_u32_e32 v56, vcc, 0x300, v13
; VI-NEXT:    v_lshlrev_b32_e32 v57, 16, v14
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v16
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x300, v17
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v54
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x300, v40
; VI-NEXT:    v_or_b32_sdwa v13, v13, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v17, v18, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v47, 16, v12
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v42
; VI-NEXT:    v_add_u32_e32 v54, vcc, 0x300, v43
; VI-NEXT:    v_or_b32_sdwa v12, v12, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v55
; VI-NEXT:    v_add_u32_e32 v55, vcc, 0x300, v41
; VI-NEXT:    v_or_b32_sdwa v10, v10, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v44
; VI-NEXT:    v_add_u32_e32 v40, vcc, 0x300, v45
; VI-NEXT:    v_or_b32_sdwa v11, v11, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v43, vcc, 0x300, v35
; VI-NEXT:    v_add_u32_e32 v35, vcc, 0x300, v36
; VI-NEXT:    v_add_u32_e32 v36, vcc, 0x300, v19
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v52
; VI-NEXT:    v_or_b32_sdwa v14, v14, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v15, v57, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v16, v47, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v19, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v20, v20, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v21, v21, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x3000000, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x3000000, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x3000000, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v6
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v7
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x3000000, v8
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x3000000, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v10
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x3000000, v11
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x3000000, v12
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x3000000, v13
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x3000000, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x3000000, v15
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x3000000, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x3000000, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x3000000, v19
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v20
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x3000000, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v54, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v55, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v31, v54, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x3000000, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v40, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v41, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v30, v40, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x3000000, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v42, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v29, v42, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x3000000, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v44, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v45, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v28, v44, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x3000000, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v46, 16, v18
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v38
; VI-NEXT:    v_or_b32_sdwa v18, v18, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v27, v46, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x3000000, v18
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x3000000, v27
; VI-NEXT:  .LBB93_3: ; %end
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB93_4:
; VI-NEXT:    v_mov_b32_e32 v43, v39
; VI-NEXT:    v_mov_b32_e32 v52, v34
; VI-NEXT:    v_mov_b32_e32 v39, v33
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v34, v53
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, v2
; VI-NEXT:    v_mov_b32_e32 v35, v1
; VI-NEXT:    v_mov_b32_e32 v55, v57
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_branch .LBB93_2
;
; GFX9-LABEL: bitcast_v128i8_to_v64f16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s63, v10
; GFX9-NEXT:    v_readfirstlane_b32 s7, v0
; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    v_readfirstlane_b32 s43, v2
; GFX9-NEXT:    v_readfirstlane_b32 s6, v1
; GFX9-NEXT:    v_readfirstlane_b32 s8, v5
; GFX9-NEXT:    v_readfirstlane_b32 s10, v4
; GFX9-NEXT:    v_readfirstlane_b32 s15, v3
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readfirstlane_b32 s12, v9
; GFX9-NEXT:    v_readfirstlane_b32 s14, v8
; GFX9-NEXT:    v_readfirstlane_b32 s47, v7
; GFX9-NEXT:    v_readfirstlane_b32 s59, v6
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; GFX9-NEXT:    v_readfirstlane_b32 s42, v13
; GFX9-NEXT:    v_readfirstlane_b32 s74, v30
; GFX9-NEXT:    v_readfirstlane_b32 s44, v29
; GFX9-NEXT:    v_readfirstlane_b32 s56, v28
; GFX9-NEXT:    v_readfirstlane_b32 s62, v27
; GFX9-NEXT:    v_readfirstlane_b32 s72, v26
; GFX9-NEXT:    v_readfirstlane_b32 s13, v25
; GFX9-NEXT:    v_readfirstlane_b32 s40, v24
; GFX9-NEXT:    v_readfirstlane_b32 s57, v23
; GFX9-NEXT:    v_readfirstlane_b32 s60, v22
; GFX9-NEXT:    v_readfirstlane_b32 s9, v21
; GFX9-NEXT:    v_readfirstlane_b32 s11, v20
; GFX9-NEXT:    v_readfirstlane_b32 s41, v19
; GFX9-NEXT:    v_readfirstlane_b32 s76, v18
; GFX9-NEXT:    v_readfirstlane_b32 s45, v17
; GFX9-NEXT:    v_readfirstlane_b32 s58, v16
; GFX9-NEXT:    v_readfirstlane_b32 s73, v15
; GFX9-NEXT:    v_readfirstlane_b32 s75, v14
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:312
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:296
; GFX9-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:292
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:280
; GFX9-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:276
; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:272
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:268
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:264
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:260
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:256
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:252
; GFX9-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:248
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:244
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:184
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:180
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:176
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:172
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:168
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:164
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:12
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    s_cbranch_scc0 .LBB93_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_mov_b32_e32 v8, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v8
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v8
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v8
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v8
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v8
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v8
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v8
; GFX9-NEXT:    v_mov_b32_e32 v4, s6
; GFX9-NEXT:    v_perm_b32 v4, s7, v4, v8
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX9-NEXT:    v_mov_b32_e32 v4, s15
; GFX9-NEXT:    v_perm_b32 v4, s43, v4, v8
; GFX9-NEXT:    v_mov_b32_e32 v5, s8
; GFX9-NEXT:    v_perm_b32 v5, s10, v5, v8
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX9-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX9-NEXT:    v_mov_b32_e32 v5, s47
; GFX9-NEXT:    v_perm_b32 v5, s59, v5, v8
; GFX9-NEXT:    v_mov_b32_e32 v6, s12
; GFX9-NEXT:    v_perm_b32 v6, s14, v6, v8
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX9-NEXT:    v_mov_b32_e32 v6, s61
; GFX9-NEXT:    v_perm_b32 v6, s63, v6, v8
; GFX9-NEXT:    v_mov_b32_e32 v7, s42
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v8
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v7, s73
; GFX9-NEXT:    v_perm_b32 v7, s75, v7, v8
; GFX9-NEXT:    v_mov_b32_e32 v63, v33
; GFX9-NEXT:    v_mov_b32_e32 v33, v9
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_perm_b32 v9, s58, v9, v8
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX9-NEXT:    v_lshl_or_b32 v7, v9, 16, v7
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v10
; GFX9-NEXT:    v_perm_b32 v10, s74, v12, v8
; GFX9-NEXT:    v_or_b32_sdwa v9, v11, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v11, v9, 16, v10
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v18
; GFX9-NEXT:    v_or_b32_sdwa v9, v22, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v13
; GFX9-NEXT:    v_or_b32_sdwa v10, v15, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_lshl_or_b32 v12, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v14
; GFX9-NEXT:    v_perm_b32 v10, v26, v21, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v17, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v13, v9, 16, v10
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
; GFX9-NEXT:    v_or_b32_sdwa v9, v30, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v16
; GFX9-NEXT:    v_or_b32_sdwa v10, v20, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v14, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v19
; GFX9-NEXT:    v_perm_b32 v10, v35, v29, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v24, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v15, v9, 16, v10
; GFX9-NEXT:    v_perm_b32 v9, v39, v34, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v51, v36, s4
; GFX9-NEXT:    v_perm_b32 v16, v28, v23, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_perm_b32 v17, v37, v27, s4
; GFX9-NEXT:    v_lshl_or_b32 v16, v16, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v41, v50, s4
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v10
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v52, v31, s4
; GFX9-NEXT:    v_lshl_or_b32 v18, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v40, v46, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v32, v38, s4
; GFX9-NEXT:    v_lshl_or_b32 v19, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v55, v57, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v45, v49, s4
; GFX9-NEXT:    v_lshl_or_b32 v20, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v53, v60, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v56, v44, s4
; GFX9-NEXT:    v_lshl_or_b32 v21, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v58, v42, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v59, v47, s4
; GFX9-NEXT:    v_lshl_or_b32 v22, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v50, v32
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v51, v49
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v41, v46
; GFX9-NEXT:    v_mov_b32_e32 v39, v52
; GFX9-NEXT:    v_mov_b32_e32 v52, v38
; GFX9-NEXT:    v_mov_b32_e32 v34, v37
; GFX9-NEXT:    v_mov_b32_e32 v37, v31
; GFX9-NEXT:    s_lshl_b32 s5, s13, 8
; GFX9-NEXT:    s_lshl_b32 s77, s9, 8
; GFX9-NEXT:    s_lshl_b32 s78, s44, 8
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_perm_b32 v9, v9, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_perm_b32 v10, v32, v58, s4
; GFX9-NEXT:    v_lshl_or_b32 v23, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v49
; GFX9-NEXT:    v_or_b32_sdwa v10, v24, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v24, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v35
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v25, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v25, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v33
; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v33, v46
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    v_or_b32_sdwa v9, v43, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v26, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v48
; GFX9-NEXT:    v_perm_b32 v10, v63, v61, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v27, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v54, v61
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_mov_b32_e32 v48, v63
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
; GFX9-NEXT:    v_or_b32_sdwa v10, v28, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshl_or_b32 v28, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v36, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v29, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v43
; GFX9-NEXT:    v_or_b32_sdwa v9, v62, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
; GFX9-NEXT:    v_or_b32_sdwa v10, v30, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshl_or_b32 v30, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
; GFX9-NEXT:    v_perm_b32 v10, v63, v61, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v46, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v31, v9, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v9, s57
; GFX9-NEXT:    v_perm_b32 v8, s60, v9, v8
; GFX9-NEXT:    s_and_b32 s4, s40, 0xff
; GFX9-NEXT:    s_or_b32 s4, s4, s5
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX9-NEXT:    v_lshl_or_b32 v9, s4, 16, v8
; GFX9-NEXT:    s_and_b32 s4, s76, 0xff
; GFX9-NEXT:    s_lshl_b32 s5, s41, 8
; GFX9-NEXT:    s_or_b32 s4, s4, s5
; GFX9-NEXT:    s_and_b32 s5, s11, 0xff
; GFX9-NEXT:    s_or_b32 s5, s5, s77
; GFX9-NEXT:    s_pack_ll_b32_b16 s4, s4, s5
; GFX9-NEXT:    s_and_b32 s5, s72, 0xff
; GFX9-NEXT:    s_lshl_b32 s77, s62, 8
; GFX9-NEXT:    s_or_b32 s5, s5, s77
; GFX9-NEXT:    s_and_b32 s77, s56, 0xff
; GFX9-NEXT:    s_or_b32 s77, s77, s78
; GFX9-NEXT:    s_pack_ll_b32_b16 s5, s5, s77
; GFX9-NEXT:    v_mov_b32_e32 v8, s4
; GFX9-NEXT:    v_mov_b32_e32 v10, s5
; GFX9-NEXT:    s_mov_b64 s[4:5], 0
; GFX9-NEXT:    v_mov_b32_e32 v63, v36
; GFX9-NEXT:    s_branch .LBB93_3
; GFX9-NEXT:  .LBB93_2:
; GFX9-NEXT:    v_mov_b32_e32 v54, v61
; GFX9-NEXT:    v_mov_b32_e32 v50, v32
; GFX9-NEXT:    v_mov_b32_e32 v51, v49
; GFX9-NEXT:    v_mov_b32_e32 v39, v52
; GFX9-NEXT:    v_mov_b32_e32 v52, v38
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v41, v46
; GFX9-NEXT:    v_mov_b32_e32 v34, v37
; GFX9-NEXT:    v_mov_b32_e32 v37, v31
; GFX9-NEXT:    s_mov_b64 s[4:5], -1
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:  .LBB93_3: ; %Flow
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_vccnz .LBB93_5
; GFX9-NEXT:  ; %bb.4: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v48
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v0, v0, v61, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v33
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v0, v0, v38, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v62
; GFX9-NEXT:    v_perm_b32 v0, v0, v43, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    s_add_i32 s74, s74, 3
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    s_add_i32 s72, s72, 3
; GFX9-NEXT:    s_add_i32 s56, s56, 3
; GFX9-NEXT:    s_add_i32 s60, s60, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s76, s76, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s75, s75, 3
; GFX9-NEXT:    s_add_i32 s58, s58, 3
; GFX9-NEXT:    s_add_i32 s63, s63, 3
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s59, s59, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    v_mov_b32_e32 v33, s17
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v63
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v38, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v62, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v54, s4
; GFX9-NEXT:    v_add_u32_e32 v63, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v30, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v35, s4
; GFX9-NEXT:    v_add_u32_e32 v29, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v36
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v28, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v49, s4
; GFX9-NEXT:    v_add_u32_e32 v27, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v26, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v32
; GFX9-NEXT:    v_perm_b32 v0, v0, v58, s4
; GFX9-NEXT:    v_add_u32_e32 v25, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v32, s23
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v42, s4
; GFX9-NEXT:    v_add_u32_e32 v24, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v59
; GFX9-NEXT:    v_perm_b32 v0, v0, v47, s4
; GFX9-NEXT:    v_add_u32_e32 v23, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v53
; GFX9-NEXT:    v_perm_b32 v0, v0, v60, s4
; GFX9-NEXT:    v_add_u32_e32 v22, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v56
; GFX9-NEXT:    v_perm_b32 v0, v0, v44, s4
; GFX9-NEXT:    v_add_u32_e32 v21, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v55
; GFX9-NEXT:    v_perm_b32 v0, v0, v57, s4
; GFX9-NEXT:    v_add_u32_e32 v20, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v45
; GFX9-NEXT:    v_perm_b32 v0, v0, v51, s4
; GFX9-NEXT:    v_add_u32_e32 v49, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v40
; GFX9-NEXT:    v_perm_b32 v0, v0, v41, s4
; GFX9-NEXT:    v_add_u32_e32 v19, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v50
; GFX9-NEXT:    v_perm_b32 v0, v0, v52, s4
; GFX9-NEXT:    v_add_u32_e32 v52, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX9-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff, v24
; GFX9-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff, v26
; GFX9-NEXT:    v_lshl_or_b32 v23, v25, 16, v23
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff, v30
; GFX9-NEXT:    v_lshl_or_b32 v25, v29, 16, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff, v28
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v24, v27, 16, v24
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff, v63
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX9-NEXT:    v_lshl_or_b32 v19, v52, 16, v19
; GFX9-NEXT:    v_lshl_or_b32 v20, v49, 16, v20
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v18, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v39
; GFX9-NEXT:    v_perm_b32 v0, v0, v37, s4
; GFX9-NEXT:    v_add_u32_e32 v37, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX9-NEXT:    v_lshl_or_b32 v18, v37, 16, v18
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v54, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v34
; GFX9-NEXT:    v_mov_b32_e32 v34, s19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v17, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v16, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v55, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v16, v55, 16, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v15, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v40, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v15, v40, 16, v15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v14, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v41, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v14, v41, 16, v14
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v13, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v42, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v13, v42, 16, v13
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v12, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v43, 0x300, v0
; GFX9-NEXT:    v_mov_b32_e32 v0, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v32, s22, v32, v0
; GFX9-NEXT:    v_add_u32_e32 v32, 0x300, v32
; GFX9-NEXT:    v_perm_b32 v33, s16, v33, v0
; GFX9-NEXT:    v_add_u32_e32 v33, 0x300, v33
; GFX9-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX9-NEXT:    v_lshl_or_b32 v12, v43, 16, v12
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, s74, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v11, 0x300, v1
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_add_u32_e32 v44, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s62
; GFX9-NEXT:    v_perm_b32 v1, s72, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v10, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s44
; GFX9-NEXT:    v_perm_b32 v1, s56, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v45, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s57
; GFX9-NEXT:    v_perm_b32 v1, s60, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v9, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s13
; GFX9-NEXT:    v_perm_b32 v1, s40, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v46, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s41
; GFX9-NEXT:    v_perm_b32 v1, s76, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v8, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s9
; GFX9-NEXT:    v_perm_b32 v1, s11, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v47, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s73
; GFX9-NEXT:    v_perm_b32 v1, s75, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v7, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s45
; GFX9-NEXT:    v_perm_b32 v1, s58, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v56, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s61
; GFX9-NEXT:    v_perm_b32 v1, s63, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v6, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s42
; GFX9-NEXT:    v_perm_b32 v1, s46, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v57, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s47
; GFX9-NEXT:    v_perm_b32 v1, s59, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v5, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s12
; GFX9-NEXT:    v_perm_b32 v1, s14, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v58, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s15
; GFX9-NEXT:    v_perm_b32 v1, s43, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v4, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    v_perm_b32 v1, s10, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v59, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s29
; GFX9-NEXT:    v_perm_b32 v1, s28, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v3, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    v_perm_b32 v1, s7, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v60, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s25
; GFX9-NEXT:    v_perm_b32 v1, s24, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s27
; GFX9-NEXT:    v_perm_b32 v1, s26, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v61, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s21
; GFX9-NEXT:    v_perm_b32 v1, s20, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX9-NEXT:    v_lshl_or_b32 v1, v32, 16, v1
; GFX9-NEXT:    v_and_b32_e32 v32, 0xffff, v54
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v32
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX9-NEXT:    v_lshl_or_b32 v29, v30, 16, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v26, v31, 16, v26
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, s18, v34, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff, v38
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v2, v61, 16, v2
; GFX9-NEXT:    v_lshl_or_b32 v3, v60, 16, v3
; GFX9-NEXT:    v_lshl_or_b32 v4, v59, 16, v4
; GFX9-NEXT:    v_lshl_or_b32 v5, v58, 16, v5
; GFX9-NEXT:    v_lshl_or_b32 v6, v57, 16, v6
; GFX9-NEXT:    v_lshl_or_b32 v7, v56, 16, v7
; GFX9-NEXT:    v_lshl_or_b32 v8, v47, 16, v8
; GFX9-NEXT:    v_lshl_or_b32 v9, v46, 16, v9
; GFX9-NEXT:    v_lshl_or_b32 v10, v45, 16, v10
; GFX9-NEXT:    v_lshl_or_b32 v11, v44, 16, v11
; GFX9-NEXT:    v_lshl_or_b32 v28, v62, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshl_or_b32 v30, v31, 16, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX9-NEXT:    v_lshl_or_b32 v31, v32, 16, v31
; GFX9-NEXT:  .LBB93_5: ; %end
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64f16_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:260
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:188
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:60
; GFX11-TRUE16-NEXT:    s_clause 0xf
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s74, v30
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v29
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v28
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v27
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v26
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v25
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v24
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v23
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v22
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v21
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v20
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v19
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v18
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v17
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v16
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s75, 0
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB93_4
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, 0xc0c0004 :: v_dual_lshlrev_b32 v13, 8, v86
; GFX11-TRUE16-NEXT:    s_and_b32 s76, s43, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s77, s11, 8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s2, s3, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s22, s23, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s26, s27, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s0, s1, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s18, s19, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s9, s4, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s24, s25, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s40, s8, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s14, s7, v8
; GFX11-TRUE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s16, s17, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s12, s6, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v4.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s28, s29, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s76 :: v_dual_lshlrev_b32 v15, 8, v70
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s61, s45, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s60, s15, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v7.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s72, s58, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v9.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s59, s44, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v68
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s41, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s89, s10, 8
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s73, s42, v8
; GFX11-TRUE16-NEXT:    s_or_b32 s88, s88, s89
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v12
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s88
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s74, 0xff
; GFX11-TRUE16-NEXT:    s_and_b32 s77, s63, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s76, s57, 8
; GFX11-TRUE16-NEXT:    s_and_b32 s78, s56, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s79, s5, 8
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, s88, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v38
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v32
; GFX11-TRUE16-NEXT:    s_or_b32 s76, s77, s76
; GFX11-TRUE16-NEXT:    s_or_b32 s77, s78, s79
; GFX11-TRUE16-NEXT:    s_and_b32 s78, s62, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s79, s47, 8
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s46, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s89, s13, 8
; GFX11-TRUE16-NEXT:    s_or_b32 s78, s78, s79
; GFX11-TRUE16-NEXT:    s_or_b32 s79, s88, s89
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v8.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v14, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v83
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v33
; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s77
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v50
; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s78, s79
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v39
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v134, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v8.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v15, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v53, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v99, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v66, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v97, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v71, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v103, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v69, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v116, v67, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v84, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v130, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v98, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v145, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v113, v55, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v135, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v147, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v118, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v117
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v64
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v151, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v164
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v119
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v131
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v82
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v149
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v8, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v176, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v26, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v144
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v87
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v128
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v167
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v101
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v160
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v102
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v8, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v28, v29
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v162
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v132
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v177, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v179
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v150
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v133
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v40, 0xff, v182
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v8, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v166
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v165
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v129
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v41, v29, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v178
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v181, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v183, v8, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s76
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v40, v31
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v180, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s77
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v30.h, v183.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v41.l
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB93_3
; GFX11-TRUE16-NEXT:  .LBB93_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v180
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v166
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v182
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v165
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v181
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s59, s59, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v160
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v179
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v162
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v1, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v177
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v102, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v149
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v132, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v167
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v144
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v176
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v101, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v131
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v164
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v87, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v117
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v151
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v118
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v82, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v147
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v113
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v85, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v135
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v55, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v98
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v145
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v55, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v130
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v52, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v69
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v116
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v71
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v67, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v103
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v66
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v97
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v53
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v99
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v50
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v38
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v134
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v114
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v68
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s74, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s73, s73, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s73, s42, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s62, s47, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s59, s44, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s43, s43, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s72, s58, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s43, s11, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s56, s5, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s12, s6, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s61, s61, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s41, s10, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s46, s13, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s60, s15, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s61, s45, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v65, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s28, s29, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s63, s63, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s63, s57, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v48, 0x300, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s9, s4, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s20, s21, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s14, s7, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v54, s40, s8, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v67, s26, s27, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v68, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s22, s23, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v69, s2, s3, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v70, s18, s19, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v66, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s24, s25, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v71, s16, s17, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v80, s0, s1, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v54, 0x300, v54
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v67, 0x300, v67
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v69
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v70
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v81, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v71
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v80
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v69.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v70.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v81.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v67.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v68.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v54.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v66.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v65.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v53.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v50.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v48.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v39.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v37.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v38.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v32.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v33.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v34.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v35.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v51.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v36.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v49.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v52.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v55.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v85.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v64.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v82.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v87.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v101.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v28.h, v132.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v102.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v30.h, v129.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v31.h, v133.l
; GFX11-TRUE16-NEXT:  .LBB93_3: ; %end
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT:  .LBB93_4:
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_branch .LBB93_2
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64f16_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:260
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:188
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:132
; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v54, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v52, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v55, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:60
; GFX11-FAKE16-NEXT:    s_clause 0xf
; GFX11-FAKE16-NEXT:    scratch_load_u16 v34, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v50, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v39, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v36, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v49, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v37, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v53, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v32, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v48, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v51, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v38, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v35, off, s32
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s74, v30
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v29
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v28
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v27
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v26
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v25
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v24
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v23
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v22
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v21
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v20
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v19
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v18
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v17
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v16
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s75, 0
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:320 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-FAKE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB93_4
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s74, 0xff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v48
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v32
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s0, s1, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s16, s17, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s20, s21, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s2, s3, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s18, s19, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s22, s23, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s24, s25, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s28, s29, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s11, s8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s6, s4, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s7, s5, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v67, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v80, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v68, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v10, 16, v11
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s26, s27, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s42, s15, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s44, s41, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v69, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v7, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s12, s9, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v116, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_lshl_b32 s77, s43, 8
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v118, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s14, s10, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, s76, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v38
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s46, 0xff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v53
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v9, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v49
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v36
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v11, 16, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v14, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v50
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v65, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v14, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v82, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v115, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v130
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v9, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v96, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v71, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v98, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v86, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v117
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v144
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v102, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s60, s57, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v87, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v135
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v150
; GFX11-FAKE16-NEXT:    s_lshl_b32 s77, s47, 8
; GFX11-FAKE16-NEXT:    s_and_b32 s78, s40, 0xff
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v11, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v112, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, s76, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v132, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s58, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s79, s13, 8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v11, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v113, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-FAKE16-NEXT:    s_or_b32 s77, s78, s79
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s73, s72, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v11, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v134
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v119
; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s77
; GFX11-FAKE16-NEXT:    s_and_b32 s77, s62, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s78, s59, 8
; GFX11-FAKE16-NEXT:    s_lshl_b32 s79, s45, 8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v24, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v131
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v114
; GFX11-FAKE16-NEXT:    s_or_b32 s77, s77, s78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_and_b32 s78, s56, 0xff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v24, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v146, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_or_b32 s78, s78, s79
; GFX11-FAKE16-NEXT:    s_and_b32 s79, s63, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s88, s61, 8
; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v160
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v147
; GFX11-FAKE16-NEXT:    s_or_b32 s78, s79, s88
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v148
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v26, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v149
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v133
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v26, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v162, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v177
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v164
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v151
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v28, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v176
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v161
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v28, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v179, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v181
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v165
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v178
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v180
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v166
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v183, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, s77
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v11, v31
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v182, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v183
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v40, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, s78, 16, v8
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, s76
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v40
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB93_3
; GFX11-FAKE16-NEXT:  .LBB93_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v182
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v180
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v178
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v179
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v176
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v177
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v3, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v4, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v5, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v151
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v150, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v151, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v162
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v149
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v148
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v132
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v135, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v160
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v131
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v144, 0x300, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v134
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v115
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v115, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v118
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v130
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v113
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v112
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v114, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v116
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v87
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v96
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v100, 0x300, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v98
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v86
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v71
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v71, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v82
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v68
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v69
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v80
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v54, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v65
; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v52, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v55, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v53
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v49
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v38
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v2
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v48
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s74, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s73, s73, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s62, s59, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s63, s63, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s73, s72, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s40, s13, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s63, s61, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s12, s9, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s56, s45, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s60, s57, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s46, s43, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s58, s47, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s11, s8, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v50, s24, s25, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, s0, s1, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s44, s41, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s14, s10, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s42, s15, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s26, s27, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v67, s2, s3, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v65, 0x300, v65
; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s7, s5, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s28, s29, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v64, s16, s17, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v67, 0x300, v67
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff, v65
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s6, s4, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s20, s21, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v66, s18, s19, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v64
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v67, 16, v65
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v69, 16, v50
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v38
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v32, 16, v36
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v34, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v54
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v17
; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s22, s23, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v66, 0x300, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v64
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v68, 0x300, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v51
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v55, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v52, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v99
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v115
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v66, 16, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v68
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v53, 16, v51
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v100, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v97, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v135
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v151
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v64
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v65
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v49, 16, v48
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v50
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v51
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v38
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v70, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v114, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v144, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v133, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v150, 16, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v35
; GFX11-FAKE16-NEXT:  .LBB93_3: ; %end
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:320 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-FAKE16-NEXT:  .LBB93_4:
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_branch .LBB93_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) {
; SI-LABEL: bitcast_v64f16_to_v128i8:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v32, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32
; SI-NEXT:    v_mov_b32_e32 v31, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v30
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v29
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v28
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v27
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v26
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v25
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v24
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v23
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v22
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v21
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v20
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v19
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v18
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v17
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v16
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v15
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v14
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v13
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v11
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v10
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v9
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v7
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v2
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v4
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v31
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v32
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v47
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v5
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v8
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v57
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; kill: killed $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB94_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v4
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_mov_b32_e32 v42, v40
; SI-NEXT:    v_mov_b32_e32 v6, v53
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v52, v2, v31
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v45
; SI-NEXT:    v_or_b32_e32 v2, v2, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v51, v31, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v43
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v50, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v48, v5, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v40
; SI-NEXT:    v_mov_b32_e32 v40, v55
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v49, v5, v31
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v38, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v55
; SI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v39, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v37, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v10
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v55
; SI-NEXT:    v_or_b32_e32 v36, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v2, v52, 24
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v11
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v34, v5, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v53
; SI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v35, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v2, v52, 16
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v13
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v32, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v53
; SI-NEXT:    v_or_b32_e32 v33, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v2, v52, 8
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v63, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v14
; SI-NEXT:    v_or_b32_e32 v31, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v50, v51, 24
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v61, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v13
; SI-NEXT:    v_or_b32_e32 v62, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v50, v51, 16
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v59, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v12
; SI-NEXT:    v_or_b32_e32 v60, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v50, v51, 8
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v20, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v11
; SI-NEXT:    v_or_b32_e32 v58, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v49, v48, 24
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v23
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v19, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v10
; SI-NEXT:    v_or_b32_e32 v56, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v49, v48, 16
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v25
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v18, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v26
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v9
; SI-NEXT:    v_or_b32_e32 v46, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v49, v48, 8
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v27
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v17, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v15
; SI-NEXT:    v_or_b32_e32 v44, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v39, v38, 24
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v29
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v1
; SI-NEXT:    v_or_b32_e32 v16, v5, v7
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v30
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v8
; SI-NEXT:    v_or_b32_e32 v41, v5, v7
; SI-NEXT:    v_alignbit_b32 v5, v39, v38, 16
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_or_b32_e32 v5, v4, v3
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v47
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v4
; SI-NEXT:    v_or_b32_e32 v54, v1, v3
; SI-NEXT:    v_alignbit_b32 v1, v39, v38, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v37, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v37, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v37, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v35, v34, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v35, v34, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v35, v34, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v32, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v32, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v32, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v31, v63, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v31, v63, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v31, v63, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v62, v61, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v62, v61, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v62, v61, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v60, v59, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v60, v59, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v60, v59, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v20, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v20, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v58, v20, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v19, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v19, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v56, v19, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v46, v18, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v46, v18, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v46, v18, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v17, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v17, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v17, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v41, v16, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v41, v16, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v41, v16, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v5, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v5, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v5, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v2
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v50
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v49
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v39
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v35
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v33
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v31
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v62
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v60
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v58
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v56
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v46
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v44
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v41
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v54
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v45, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v43, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v42, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v40, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v55, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v6, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v53, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v14, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v13, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v12, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v11, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v10, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v9, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v15, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v8, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    v_mov_b32_e32 v45, v17
; SI-NEXT:    v_mov_b32_e32 v43, v16
; SI-NEXT:    v_mov_b32_e32 v55, v40
; SI-NEXT:    v_mov_b32_e32 v40, v42
; SI-NEXT:    v_mov_b32_e32 v42, v5
; SI-NEXT:    v_mov_b32_e32 v53, v6
; SI-NEXT:    v_bfe_u32 v6, v4, 8, 8
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:  .LBB94_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB94_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v57
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v47
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v42, v3, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
; SI-NEXT:    v_or_b32_e32 v54, v1, v2
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v29
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v43, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v29
; SI-NEXT:    v_or_b32_e32 v41, v1, v2
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v27
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v45, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    v_or_b32_e32 v44, v1, v2
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v25
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v28, v2, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v25
; SI-NEXT:    v_or_b32_e32 v46, v1, v2
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v23
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v26, v2, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v2
; SI-NEXT:    v_mov_b32_e32 v24, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v21
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v23
; SI-NEXT:    v_or_b32_e32 v56, v1, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v22
; SI-NEXT:    v_or_b32_e32 v22, v3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v21
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v17
; SI-NEXT:    v_or_b32_e32 v58, v2, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v19
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v59, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v20
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v19
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_or_b32_e32 v60, v3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v61, v4, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v18
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v15
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v17
; SI-NEXT:    v_or_b32_e32 v62, v1, v3
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v16
; SI-NEXT:    v_or_b32_e32 v63, v3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v11
; SI-NEXT:    v_or_b32_e32 v31, v2, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v13
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v32, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v14
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v13
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_or_b32_e32 v33, v3, v1
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v53
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v34, v4, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v12
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v53, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v9
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v53
; SI-NEXT:    v_or_b32_e32 v35, v1, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v10
; SI-NEXT:    v_or_b32_e32 v37, v3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v5
; SI-NEXT:    v_or_b32_e32 v36, v2, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v7
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v55
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v55, v3
; SI-NEXT:    v_or_b32_e32 v38, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v8
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v55
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_or_b32_e32 v39, v3, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v40
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v40, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v48, v4, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v40
; SI-NEXT:    v_or_b32_e32 v49, v1, v3
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_or_b32_e32 v51, v3, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v24
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v5
; SI-NEXT:    v_or_b32_e32 v50, v2, v3
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_or_b32_e32 v52, v2, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v7
; SI-NEXT:    v_or_b32_e32 v2, v3, v1
; SI-NEXT:    v_alignbit_b32 v1, v2, v52, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v2, v52, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v2, v52, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v50, v51, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v50, v51, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v50, v51, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v49, v48, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v49, v48, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v49, v48, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v39, v38, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v39, v38, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v39, v38, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v37, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v37, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v37, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v35, v34, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v35, v34, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v35, v34, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v32, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v32, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v32, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v31, v63, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v31, v63, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v31, v63, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v62, v61, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v62, v61, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v62, v61, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v60, v59, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v60, v59, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v60, v59, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v22, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v22, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v58, v22, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v26, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v26, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v56, v26, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v46, v28, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v46, v28, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v46, v28, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v45, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v45, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v45, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v41, v43, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v41, v43, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v41, v43, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v42, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v42, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v42, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v2
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v50
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v49
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v39
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v35
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v33
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v31
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v62
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v60
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v58
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v56
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v46
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v44
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v41
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v54
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v7, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v5, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v40, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v55, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v9, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v53, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v13, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v15, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v17, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v19, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v21, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v23, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v25, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v27, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_bfe_u32 v1, v29, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_bfe_u32 v6, v6, 8, 8
; SI-NEXT:  .LBB94_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v52
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
; SI-NEXT:    v_or_b32_e32 v1, v1, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_or_b32_e32 v1, v1, v3
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v51
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v50
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v48
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v49
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v40
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v38
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v39
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v55
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v37
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v36
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v34
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v35
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v53
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v33
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v63
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v31
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v61
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v62
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v59
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v60
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v58
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v56
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v46
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v45
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v44
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v41
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v42
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v54
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v6
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v128i8:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v41, 16, v24
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v31, v30
; VI-NEXT:    v_mov_b32_e32 v30, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v16
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v11
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v9
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v30
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    v_lshrrev_b32_e32 v40, 16, v14
; VI-NEXT:    v_lshrrev_b32_e32 v45, 16, v12
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v58, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v52, 16, v31
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v28
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v42, 16, v26
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v25
; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v22
; VI-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
; VI-NEXT:    v_lshrrev_b32_e32 v49, 16, v20
; VI-NEXT:    v_lshrrev_b32_e32 v56, 16, v19
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v18
; VI-NEXT:    v_lshrrev_b32_e32 v63, 16, v17
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr24
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; kill: killed $vgpr24
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    v_lshrrev_b32_e32 v36, 16, v23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v29
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshrrev_b32_e32 v44, 16, v51
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshrrev_b32_e32 v61, 16, v50
; VI-NEXT:    ; kill: killed $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB94_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v16
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v16
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v15
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v14
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v14
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v13
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v12
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v12
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v11
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v10
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v10
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v9
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v8
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v8
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v7
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v6
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v6
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v5
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v4
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v4
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v3
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v2
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v2
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v1
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v51
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v51
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v50
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v31
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v31
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v30
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v28
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v28
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v27
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v26
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v26
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v25
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v24, 8, v20
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v24, 8, v19
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v29, v49
; VI-NEXT:    v_lshrrev_b64 v[48:49], 24, v[5:6]
; VI-NEXT:    v_mov_b32_e32 v49, v29
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_lshrrev_b32_e32 v23, 24, v55
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v55
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v54
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v22
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v21
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v18
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v23, 8, v17
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[7:8]
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[50:51]
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[30:31]
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v29, v44
; VI-NEXT:    v_lshrrev_b64 v[43:44], 24, v[3:4]
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[27:28]
; VI-NEXT:    v_mov_b32_e32 v44, v29
; VI-NEXT:    v_mov_b32_e32 v29, v47
; VI-NEXT:    v_lshrrev_b64 v[46:47], 24, v[1:2]
; VI-NEXT:    v_mov_b32_e32 v47, v29
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v29, v63
; VI-NEXT:    v_lshrrev_b64 v[62:63], 24, v[54:55]
; VI-NEXT:    v_lshrrev_b64 v[23:24], 24, v[25:26]
; VI-NEXT:    v_mov_b32_e32 v63, v29
; VI-NEXT:    v_mov_b32_e32 v29, v61
; VI-NEXT:    v_lshrrev_b64 v[60:61], 24, v[21:22]
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[19:20]
; VI-NEXT:    v_mov_b32_e32 v61, v29
; VI-NEXT:    v_mov_b32_e32 v29, v56
; VI-NEXT:    v_lshrrev_b64 v[55:56], 24, v[17:18]
; VI-NEXT:    v_lshrrev_b32_e32 v33, 24, v22
; VI-NEXT:    v_lshrrev_b32_e32 v57, 24, v20
; VI-NEXT:    v_lshrrev_b32_e32 v38, 24, v18
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v56, v29
; VI-NEXT:  .LBB94_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB94_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v42, 0x200
; VI-NEXT:    v_add_f16_sdwa v23, v18, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v23
; VI-NEXT:    v_add_f16_e32 v18, 0x200, v18
; VI-NEXT:    v_add_f16_sdwa v23, v17, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v57, v18, v29
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v23
; VI-NEXT:    v_add_f16_e32 v17, 0x200, v17
; VI-NEXT:    v_add_f16_sdwa v23, v20, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v56, v17, v29
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v23
; VI-NEXT:    v_add_f16_e32 v20, 0x200, v20
; VI-NEXT:    v_add_f16_sdwa v23, v19, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v19, 0x200, v19
; VI-NEXT:    v_or_b32_e32 v59, v20, v29
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v23
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    v_add_f16_sdwa v20, v51, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v51, 0x200, v51
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    v_add_f16_sdwa v37, v10, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v10, 0x200, v10
; VI-NEXT:    v_add_f16_sdwa v45, v14, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v14, 0x200, v14
; VI-NEXT:    v_or_b32_e32 v58, v19, v29
; VI-NEXT:    v_add_f16_sdwa v19, v22, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_e32 v22, 0x200, v22
; VI-NEXT:    v_add_f16_sdwa v19, v21, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v21, 0x200, v21
; VI-NEXT:    v_or_b32_e32 v55, v22, v29
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_add_f16_sdwa v19, v24, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    v_add_f16_e32 v24, 0x200, v24
; VI-NEXT:    v_add_f16_sdwa v22, v8, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v8, 0x200, v8
; VI-NEXT:    v_or_b32_e32 v54, v21, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_sdwa v19, v23, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v23, 0x200, v23
; VI-NEXT:    v_or_b32_e32 v61, v24, v29
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_sdwa v19, v26, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    v_add_f16_e32 v26, 0x200, v26
; VI-NEXT:    v_add_f16_sdwa v24, v4, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v4, 0x200, v4
; VI-NEXT:    v_add_f16_sdwa v21, v6, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v6, 0x200, v6
; VI-NEXT:    v_or_b32_e32 v60, v23, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_sdwa v19, v25, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v25, 0x200, v25
; VI-NEXT:    v_or_b32_e32 v63, v26, v29
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_sdwa v19, v28, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    v_add_f16_e32 v28, 0x200, v28
; VI-NEXT:    v_add_f16_sdwa v26, v31, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v31, 0x200, v31
; VI-NEXT:    v_add_f16_sdwa v23, v12, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v12, 0x200, v12
; VI-NEXT:    v_or_b32_e32 v62, v25, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_sdwa v19, v27, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v27, 0x200, v27
; VI-NEXT:    v_or_b32_e32 v39, v28, v29
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_sdwa v19, v50, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v50, 0x200, v50
; VI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    v_add_f16_sdwa v25, v30, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v30, 0x200, v30
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v38, v27, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v26
; VI-NEXT:    v_or_b32_e32 v34, v31, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v25
; VI-NEXT:    v_or_b32_e32 v33, v30, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    v_or_b32_e32 v36, v51, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_sdwa v19, v2, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v35, v50, v29
; VI-NEXT:    v_add_f16_e32 v2, 0x200, v2
; VI-NEXT:    v_add_f16_sdwa v20, v1, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v19
; VI-NEXT:    v_add_f16_e32 v1, 0x200, v1
; VI-NEXT:    v_or_b32_e32 v47, v2, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v46, v1, v29
; VI-NEXT:    v_add_f16_sdwa v20, v3, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v24
; VI-NEXT:    v_add_f16_e32 v3, 0x200, v3
; VI-NEXT:    v_or_b32_e32 v44, v4, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v43, v3, v29
; VI-NEXT:    v_add_f16_sdwa v20, v5, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v21
; VI-NEXT:    v_add_f16_e32 v5, 0x200, v5
; VI-NEXT:    v_or_b32_e32 v49, v6, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v48, v5, v29
; VI-NEXT:    v_add_f16_sdwa v20, v7, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v22
; VI-NEXT:    v_add_f16_e32 v7, 0x200, v7
; VI-NEXT:    v_or_b32_e32 v51, v8, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v50, v7, v29
; VI-NEXT:    v_add_f16_sdwa v20, v9, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v37
; VI-NEXT:    v_add_f16_e32 v9, 0x200, v9
; VI-NEXT:    v_or_b32_e32 v53, v10, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v52, v9, v29
; VI-NEXT:    v_add_f16_sdwa v20, v11, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v23
; VI-NEXT:    v_add_f16_e32 v11, 0x200, v11
; VI-NEXT:    v_or_b32_e32 v41, v12, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v40, v11, v29
; VI-NEXT:    v_add_f16_sdwa v20, v13, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v45
; VI-NEXT:    v_add_f16_e32 v13, 0x200, v13
; VI-NEXT:    v_or_b32_e32 v32, v14, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v20
; VI-NEXT:    v_mov_b32_e32 v27, v30
; VI-NEXT:    v_add_f16_sdwa v25, v16, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_mov_b32_e32 v28, v31
; VI-NEXT:    v_or_b32_e32 v31, v13, v29
; VI-NEXT:    v_add_f16_e32 v16, 0x200, v16
; VI-NEXT:    v_add_f16_sdwa v42, v15, v42 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v25
; VI-NEXT:    v_add_f16_e32 v15, 0x200, v15
; VI-NEXT:    v_or_b32_e32 v30, v16, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v42
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v29, v15, v29
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v30
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v29
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v32
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v31
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v41
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v40
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v53
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v52
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v51
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v50
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v49
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v48
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v44
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v43
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v47
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v46
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v36
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v35
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v34
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v33
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v39
; VI-NEXT:    v_lshrrev_b64 v[29:30], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v38
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v63
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v62
; VI-NEXT:    v_lshrrev_b64 v[29:30], 24, v[31:32]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v61
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v60
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v55
; VI-NEXT:    v_mov_b32_e32 v31, v28
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v54
; VI-NEXT:    v_mov_b32_e32 v30, v27
; VI-NEXT:    v_lshrrev_b64 v[27:28], 24, v[40:41]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v59
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[33:34]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v58
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v57
; VI-NEXT:    v_lshrrev_b64 v[27:28], 24, v[52:53]
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v20, 8, v56
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[38:39]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v20, v25, 8, 8
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[43:44], 24, v[43:44]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v40, v45
; VI-NEXT:    v_bfe_u32 v20, v45, 8, 8
; VI-NEXT:    v_mov_b32_e32 v45, v23
; VI-NEXT:    v_mov_b32_e32 v52, v26
; VI-NEXT:    v_lshrrev_b64 v[26:27], 24, v[50:51]
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v20, v45, 8, 8
; VI-NEXT:    v_lshrrev_b64 v[35:36], 24, v[35:36]
; VI-NEXT:    v_lshrrev_b64 v[33:34], 24, v[62:63]
; VI-NEXT:    v_lshrrev_b64 v[62:63], 24, v[60:61]
; VI-NEXT:    v_lshrrev_b64 v[60:61], 24, v[54:55]
; VI-NEXT:    v_lshrrev_b64 v[54:55], 24, v[58:59]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v59, v37
; VI-NEXT:    v_bfe_u32 v20, v37, 8, 8
; VI-NEXT:    v_mov_b32_e32 v37, v22
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v20, v37, 8, 8
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    v_lshrrev_b64 v[48:49], 24, v[48:49]
; VI-NEXT:    v_lshrrev_b64 v[46:47], 24, v[46:47]
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v20, v21, 8, 8
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    v_lshrrev_b64 v[55:56], 24, v[56:57]
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v20, v24, 8, 8
; VI-NEXT:    v_mov_b32_e32 v35, v19
; VI-NEXT:    v_bfe_u32 v19, v19, 8, 8
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v58, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v34, v24
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_bfe_u32 v23, v44, 8, 8
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v23, v52, 8, 8
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v23, v32, 8, 8
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v23, v42, 8, 8
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v33, v53, 8, 8
; VI-NEXT:    v_bfe_u32 v23, v41, 8, 8
; VI-NEXT:    v_bfe_u32 v57, v49, 8, 8
; VI-NEXT:    v_bfe_u32 v38, v47, 8, 8
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:  .LBB94_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v24, v46, s4
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v24, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v29
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v2, v2, v24, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v35, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v43, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v4, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v34, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v48, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v5, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v6, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v58, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v7, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v8, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v37, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v9, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v10, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v59, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v11, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v12, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v45, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v13, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v14, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v40, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v15, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v16, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v63, v55, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v17, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v2, v47, v38, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v18, v1, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v1, v56, v54, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v19, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v2, v49, v57, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v20, v1, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v2, v39, v60, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v21, v1, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v2, v53, v33, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v22, v1, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    v_perm_b32 v2, v36, v62, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v3, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v41, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v25, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v26, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v42, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v27, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v28, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v32, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_perm_b32 v1, v30, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v31, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v52, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v61, v2, s4
; VI-NEXT:    v_perm_b32 v1, v50, v1, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v51, v1, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v44, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v128i8:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(29)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB94_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(38)
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v31
; GFX9-NEXT:    v_lshrrev_b64 v[48:49], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v30
; GFX9-NEXT:    v_lshrrev_b64 v[49:50], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; GFX9-NEXT:    v_lshrrev_b64 v[50:51], 24, v[11:12]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v30
; GFX9-NEXT:    v_lshrrev_b64 v[51:52], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[3:4]
; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 24, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; GFX9-NEXT:    v_lshrrev_b64 v[52:53], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b64 v[42:43], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v29
; GFX9-NEXT:    v_mov_b32_e32 v43, v54
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v28
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v28
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v27
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v26
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v26
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v24
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v23
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v22
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v22
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[19:20]
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v3
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v1
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 16, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 24, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 8, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v17
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; GFX9-NEXT:  .LBB94_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB94_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    s_movk_i32 s6, 0x200
; GFX9-NEXT:    v_pk_add_f16 v32, v32, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(28)
; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX9-NEXT:    v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; GFX9-NEXT:    v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; GFX9-NEXT:    v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; GFX9-NEXT:    v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; GFX9-NEXT:    v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; GFX9-NEXT:    v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v32
; GFX9-NEXT:    v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v31
; GFX9-NEXT:    v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; GFX9-NEXT:    v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; GFX9-NEXT:    v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v28
; GFX9-NEXT:    v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; GFX9-NEXT:    v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v26
; GFX9-NEXT:    v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; GFX9-NEXT:    v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v24
; GFX9-NEXT:    v_lshrrev_b64 v[48:49], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
; GFX9-NEXT:    v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[49:50], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v24
; GFX9-NEXT:    v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[50:51], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; GFX9-NEXT:    v_lshrrev_b64 v[51:52], 24, v[9:10]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v23
; GFX9-NEXT:    v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[52:53], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v22
; GFX9-NEXT:    v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[3:4]
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v22
; GFX9-NEXT:    v_lshrrev_b64 v[42:43], 24, v[1:2]
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v22
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v16
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v3
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v1
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 16, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 24, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 8, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v17
; GFX9-NEXT:  .LBB94_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v1, v1, v33, s4
; GFX9-NEXT:    v_perm_b32 v33, v62, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v55, v63, v42, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX9-NEXT:    v_perm_b32 v39, v39, v53, s4
; GFX9-NEXT:    v_perm_b32 v2, v2, v45, s4
; GFX9-NEXT:    v_perm_b32 v3, v3, v44, s4
; GFX9-NEXT:    v_perm_b32 v4, v4, v61, s4
; GFX9-NEXT:    v_perm_b32 v5, v5, v60, s4
; GFX9-NEXT:    v_perm_b32 v37, v37, v54, s4
; GFX9-NEXT:    v_perm_b32 v17, v17, v59, s4
; GFX9-NEXT:    v_perm_b32 v18, v18, v58, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v40, v41, v40, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v52, v41, v52, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v7, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v51, v41, v51, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v9, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v10, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v50, v41, v50, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v11, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v12, v12, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v49, v41, v49, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v14, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v48, v41, v48, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v15, v15, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v41, s4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v53, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshl_or_b32 v1, v33, 16, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshl_or_b32 v1, v40, 16, v5
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v6
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshl_or_b32 v1, v52, 16, v7
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v8
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshl_or_b32 v1, v51, 16, v9
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v10
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshl_or_b32 v1, v50, 16, v11
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v12
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshl_or_b32 v1, v49, 16, v13
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v14
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshl_or_b32 v1, v48, 16, v15
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v2, v35, v34, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v43, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v16
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshl_or_b32 v1, v37, 16, v17
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    v_perm_b32 v1, v38, v36, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v18
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    v_perm_b32 v1, v19, v57, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    v_perm_b32 v1, v20, v56, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v1, v21, v47, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v46, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v22, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v23, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v24, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v25, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v26, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v27, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v28, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v30, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v31, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v32, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64f16_to_v128i8:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x2
; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB94_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 8, v17
; GFX11-TRUE16-NEXT:  .LBB94_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB94_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v32, 0x200, v32 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 8, v17
; GFX11-TRUE16-NEXT:  .LBB94_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v1.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v2.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v3.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v4.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v165, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v50, v50, v164, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v3, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v53, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v5.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v50, 16, v2
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v53, 16, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v6.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v7.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v55, 16, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v8.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v50, v50, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v53, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v39, 16, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v9.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v50, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v53, 16, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v10.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v11.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v55, 16, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v50, v50, v117, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v51, v53, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v12.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v13.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v14.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v15.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v50, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v51, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v12, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v13, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v14, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v15, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v16.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v17.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v18.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v19.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v16, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v17, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v66, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v18, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v19, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v20.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v21.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v22.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v23.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v20, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v147, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v21, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v22, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v23, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v14, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v24.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v15, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v25.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v26.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v27.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v24, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v25, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v26, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v27, v103, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v28.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v29.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v30.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v31.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v28, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v99, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v29, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v30, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v31, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v32, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v29
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64f16_to_v128i8:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x2
; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_clause 0x14 ; 84-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:12
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB94_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v1
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v32
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[17:18]
; GFX11-FAKE16-NEXT:  .LBB94_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB94_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v32, 0x200, v32 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v17
; GFX11-FAKE16-NEXT:  .LBB94_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v161, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v51, v149, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v134, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v132, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v116, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v51, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v48, v129, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v51, v119, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v114, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v37, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v103, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v101, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v98, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v96, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v48, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v51, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v52, 16, v8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v36, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v37, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v35, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v13, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v85, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v14, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v83, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v15, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v80, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v16, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v70, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v17, v76, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v75, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v18, v74, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v73, v72, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v19, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v62, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v20, v61, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v60, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v21, v58, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v57, v65, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v47, v46, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v23, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v44, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v24, v43, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v42, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v25, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v183, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v26, v182, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v181, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v178, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v28, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v176, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v29, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v165, v49, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v30, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v163, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v31, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v148, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v146, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v144, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x14 ; 84-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:92
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define inreg <128 x i8> @bitcast_v64f16_to_v128i8_scalar(<64 x half> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64f16_to_v128i8_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_writelane_b32 v34, s30, 0
; SI-NEXT:    v_writelane_b32 v34, s31, 1
; SI-NEXT:    v_writelane_b32 v34, s34, 2
; SI-NEXT:    v_writelane_b32 v34, s35, 3
; SI-NEXT:    v_writelane_b32 v34, s36, 4
; SI-NEXT:    v_writelane_b32 v34, s37, 5
; SI-NEXT:    v_writelane_b32 v34, s38, 6
; SI-NEXT:    v_writelane_b32 v34, s39, 7
; SI-NEXT:    v_writelane_b32 v34, s48, 8
; SI-NEXT:    v_writelane_b32 v34, s49, 9
; SI-NEXT:    v_writelane_b32 v34, s50, 10
; SI-NEXT:    v_writelane_b32 v34, s51, 11
; SI-NEXT:    v_writelane_b32 v34, s52, 12
; SI-NEXT:    v_writelane_b32 v34, s53, 13
; SI-NEXT:    v_writelane_b32 v34, s54, 14
; SI-NEXT:    v_writelane_b32 v34, s55, 15
; SI-NEXT:    v_writelane_b32 v34, s64, 16
; SI-NEXT:    v_writelane_b32 v34, s65, 17
; SI-NEXT:    v_writelane_b32 v34, s66, 18
; SI-NEXT:    v_writelane_b32 v34, s67, 19
; SI-NEXT:    v_writelane_b32 v34, s68, 20
; SI-NEXT:    v_writelane_b32 v34, s69, 21
; SI-NEXT:    v_writelane_b32 v34, s70, 22
; SI-NEXT:    v_writelane_b32 v34, s71, 23
; SI-NEXT:    v_writelane_b32 v34, s80, 24
; SI-NEXT:    v_writelane_b32 v34, s81, 25
; SI-NEXT:    v_writelane_b32 v34, s82, 26
; SI-NEXT:    s_lshr_b32 s5, s16, 16
; SI-NEXT:    ; implicit-def: $vgpr37 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v34, s83, 27
; SI-NEXT:    s_lshr_b32 s6, s18, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v37, s5, 0
; SI-NEXT:    v_writelane_b32 v34, s84, 28
; SI-NEXT:    s_lshr_b32 s7, s20, 16
; SI-NEXT:    v_writelane_b32 v37, s6, 2
; SI-NEXT:    v_writelane_b32 v34, s85, 29
; SI-NEXT:    s_lshr_b32 s8, s22, 16
; SI-NEXT:    v_writelane_b32 v37, s7, 4
; SI-NEXT:    v_writelane_b32 v34, s86, 30
; SI-NEXT:    s_lshr_b32 s9, s24, 16
; SI-NEXT:    v_writelane_b32 v37, s8, 5
; SI-NEXT:    v_writelane_b32 v34, s87, 31
; SI-NEXT:    s_lshr_b32 s10, s26, 16
; SI-NEXT:    v_writelane_b32 v37, s9, 6
; SI-NEXT:    v_writelane_b32 v34, s96, 32
; SI-NEXT:    s_lshr_b32 s11, s28, 16
; SI-NEXT:    v_readfirstlane_b32 s31, v3
; SI-NEXT:    v_writelane_b32 v37, s10, 7
; SI-NEXT:    v_writelane_b32 v34, s97, 33
; SI-NEXT:    v_readfirstlane_b32 s37, v5
; SI-NEXT:    s_lshr_b32 s12, s31, 16
; SI-NEXT:    v_writelane_b32 v37, s11, 8
; SI-NEXT:    v_writelane_b32 v34, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s81, v18
; SI-NEXT:    v_readfirstlane_b32 s83, v17
; SI-NEXT:    v_readfirstlane_b32 s70, v16
; SI-NEXT:    v_readfirstlane_b32 s34, v15
; SI-NEXT:    v_readfirstlane_b32 s66, v14
; SI-NEXT:    v_readfirstlane_b32 s68, v13
; SI-NEXT:    v_readfirstlane_b32 s54, v12
; SI-NEXT:    v_readfirstlane_b32 s64, v11
; SI-NEXT:    v_readfirstlane_b32 s51, v10
; SI-NEXT:    v_readfirstlane_b32 s52, v9
; SI-NEXT:    v_readfirstlane_b32 s39, v8
; SI-NEXT:    v_readfirstlane_b32 s49, v7
; SI-NEXT:    v_readfirstlane_b32 s35, v6
; SI-NEXT:    s_lshr_b32 s13, s37, 16
; SI-NEXT:    v_readfirstlane_b32 s95, v4
; SI-NEXT:    v_readfirstlane_b32 s91, v2
; SI-NEXT:    v_readfirstlane_b32 s93, v1
; SI-NEXT:    v_writelane_b32 v37, s12, 9
; SI-NEXT:    v_writelane_b32 v34, s99, 35
; SI-NEXT:    s_mov_b32 s53, s16
; SI-NEXT:    s_lshr_b32 s90, s29, 16
; SI-NEXT:    s_lshr_b32 s89, s27, 16
; SI-NEXT:    s_lshr_b32 s88, s25, 16
; SI-NEXT:    s_lshr_b32 s79, s23, 16
; SI-NEXT:    s_lshr_b32 s78, s21, 16
; SI-NEXT:    s_lshr_b32 s77, s19, 16
; SI-NEXT:    s_lshr_b32 s76, s17, 16
; SI-NEXT:    s_lshr_b32 s82, s81, 16
; SI-NEXT:    s_lshr_b32 s84, s83, 16
; SI-NEXT:    s_lshr_b32 s71, s70, 16
; SI-NEXT:    s_lshr_b32 s80, s34, 16
; SI-NEXT:    s_lshr_b32 s67, s66, 16
; SI-NEXT:    s_lshr_b32 s69, s68, 16
; SI-NEXT:    s_lshr_b32 s55, s54, 16
; SI-NEXT:    s_lshr_b32 s65, s64, 16
; SI-NEXT:    s_lshr_b32 s30, s51, 16
; SI-NEXT:    s_lshr_b32 s16, s52, 16
; SI-NEXT:    s_lshr_b32 s48, s39, 16
; SI-NEXT:    s_lshr_b32 s14, s49, 16
; SI-NEXT:    s_lshr_b32 s36, s35, 16
; SI-NEXT:    s_lshr_b32 s94, s95, 16
; SI-NEXT:    s_lshr_b32 s92, s91, 16
; SI-NEXT:    s_lshr_b32 s99, s93, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v19
; SI-NEXT:    v_writelane_b32 v37, s13, 10
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v37, s14, 11
; SI-NEXT:    v_writelane_b32 v37, s16, 12
; SI-NEXT:    ; implicit-def: $vgpr36 : SGPR spill to VGPR lane
; SI-NEXT:    ; implicit-def: $vgpr35 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB95_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s53, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s74, s4, s5
; SI-NEXT:    s_and_b32 s4, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s76, 16
; SI-NEXT:    s_or_b32 s75, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_or_b32 s72, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s77, 16
; SI-NEXT:    s_or_b32 s73, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s7, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s79, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s9, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s10, 16
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s89, 16
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s11, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s93, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s91, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s31, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s12, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s95, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s94, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s37, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s13, 16
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s35, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s36, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s49, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s14, 16
; SI-NEXT:    s_or_b32 s14, s4, s5
; SI-NEXT:    s_and_b32 s4, s39, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s48, 16
; SI-NEXT:    s_or_b32 s15, s4, s5
; SI-NEXT:    s_and_b32 s4, s52, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s16, 16
; SI-NEXT:    s_or_b32 s12, s4, s5
; SI-NEXT:    s_and_b32 s4, s51, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s13, s4, s5
; SI-NEXT:    s_and_b32 s4, s64, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s65, 16
; SI-NEXT:    s_or_b32 s10, s4, s5
; SI-NEXT:    s_and_b32 s4, s54, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s55, 16
; SI-NEXT:    s_or_b32 s11, s4, s5
; SI-NEXT:    s_and_b32 s4, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s69, 16
; SI-NEXT:    s_or_b32 s8, s4, s5
; SI-NEXT:    s_and_b32 s4, s66, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s67, 16
; SI-NEXT:    s_lshr_b32 s16, s75, 8
; SI-NEXT:    s_or_b32 s9, s4, s5
; SI-NEXT:    s_and_b32 s4, s34, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s80, 16
; SI-NEXT:    v_writelane_b32 v35, s16, 1
; SI-NEXT:    s_lshr_b32 s16, s73, 8
; SI-NEXT:    s_or_b32 s6, s4, s5
; SI-NEXT:    s_and_b32 s4, s70, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    v_writelane_b32 v35, s16, 0
; SI-NEXT:    s_lshr_b32 s16, s63, 8
; SI-NEXT:    s_or_b32 s7, s4, s5
; SI-NEXT:    s_and_b32 s4, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s84, 16
; SI-NEXT:    v_writelane_b32 v36, s16, 63
; SI-NEXT:    s_lshr_b32 s16, s59, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s81, 0xffff
; SI-NEXT:    s_lshl_b32 vcc_lo, s82, 16
; SI-NEXT:    v_writelane_b32 v36, s16, 62
; SI-NEXT:    s_lshr_b32 s16, s57, 8
; SI-NEXT:    s_or_b32 s5, s5, vcc_lo
; SI-NEXT:    v_writelane_b32 v36, s16, 61
; SI-NEXT:    s_lshr_b32 vcc_lo, s61, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 60
; SI-NEXT:    s_lshr_b32 vcc_lo, s47, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 59
; SI-NEXT:    s_lshr_b32 vcc_lo, s45, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 58
; SI-NEXT:    s_lshr_b32 vcc_lo, s43, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 57
; SI-NEXT:    s_lshr_b32 vcc_lo, s41, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 56
; SI-NEXT:    s_lshr_b32 vcc_lo, s15, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 55
; SI-NEXT:    s_lshr_b32 vcc_lo, s13, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 54
; SI-NEXT:    s_lshr_b32 vcc_lo, s11, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 53
; SI-NEXT:    s_lshr_b32 vcc_lo, s9, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 52
; SI-NEXT:    s_lshr_b32 vcc_lo, s7, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 51
; SI-NEXT:    s_lshr_b32 vcc_lo, s5, 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 50
; SI-NEXT:    s_lshr_b64 vcc, s[74:75], 24
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 17
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 18
; SI-NEXT:    s_lshr_b64 vcc, s[74:75], 16
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 15
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 16
; SI-NEXT:    s_lshr_b64 vcc, s[74:75], 8
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 13
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 14
; SI-NEXT:    s_lshr_b64 vcc, s[72:73], 24
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 23
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 24
; SI-NEXT:    s_lshr_b64 vcc, s[72:73], 16
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 21
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 22
; SI-NEXT:    s_lshr_b64 vcc, s[72:73], 8
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 19
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 20
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 24
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 29
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 30
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 16
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 27
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 28
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 8
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 25
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 26
; SI-NEXT:    s_lshr_b64 vcc, s[58:59], 24
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 35
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 36
; SI-NEXT:    s_lshr_b64 vcc, s[58:59], 16
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 33
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 34
; SI-NEXT:    s_lshr_b64 vcc, s[58:59], 8
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 31
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 32
; SI-NEXT:    s_lshr_b64 vcc, s[56:57], 24
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 41
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 42
; SI-NEXT:    s_lshr_b64 vcc, s[56:57], 16
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 39
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 40
; SI-NEXT:    s_lshr_b64 vcc, s[56:57], 8
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 37
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 38
; SI-NEXT:    s_lshr_b64 vcc, s[60:61], 24
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 47
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 48
; SI-NEXT:    s_lshr_b64 vcc, s[60:61], 16
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 45
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 46
; SI-NEXT:    s_lshr_b64 vcc, s[60:61], 8
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 43
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 44
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 24
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 53
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 54
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 16
; SI-NEXT:    s_bfe_u32 s16, s76, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 51
; SI-NEXT:    v_writelane_b32 v36, s16, 49
; SI-NEXT:    s_bfe_u32 s16, s77, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 52
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 8
; SI-NEXT:    v_writelane_b32 v36, s16, 48
; SI-NEXT:    s_bfe_u32 s16, s78, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 49
; SI-NEXT:    v_writelane_b32 v36, s16, 47
; SI-NEXT:    s_bfe_u32 s16, s79, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 50
; SI-NEXT:    s_lshr_b64 vcc, s[44:45], 24
; SI-NEXT:    v_writelane_b32 v36, s16, 46
; SI-NEXT:    s_bfe_u32 s16, s88, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 59
; SI-NEXT:    v_writelane_b32 v36, s16, 45
; SI-NEXT:    s_bfe_u32 s16, s89, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 60
; SI-NEXT:    s_lshr_b64 vcc, s[44:45], 16
; SI-NEXT:    v_writelane_b32 v36, s16, 44
; SI-NEXT:    s_bfe_u32 s16, s90, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 57
; SI-NEXT:    v_writelane_b32 v36, s16, 43
; SI-NEXT:    s_bfe_u32 s16, s92, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 58
; SI-NEXT:    s_lshr_b64 vcc, s[44:45], 8
; SI-NEXT:    v_writelane_b32 v36, s16, 42
; SI-NEXT:    s_bfe_u32 s16, s94, 0x80008
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 55
; SI-NEXT:    v_writelane_b32 v36, s16, 41
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 56
; SI-NEXT:    s_lshr_b64 vcc, s[42:43], 24
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 1
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 2
; SI-NEXT:    s_lshr_b64 vcc, s[42:43], 16
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 63
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 0
; SI-NEXT:    s_lshr_b64 vcc, s[42:43], 8
; SI-NEXT:    v_writelane_b32 v37, vcc_lo, 61
; SI-NEXT:    v_writelane_b32 v37, vcc_hi, 62
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 24
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 7
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 8
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 16
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 5
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 6
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 3
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 4
; SI-NEXT:    s_lshr_b64 vcc, s[14:15], 24
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 13
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 14
; SI-NEXT:    s_lshr_b64 vcc, s[14:15], 16
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 11
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 12
; SI-NEXT:    s_lshr_b64 vcc, s[14:15], 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 9
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 10
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 24
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 19
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 20
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 16
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 17
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 18
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 15
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 16
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 24
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 25
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 26
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 16
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 23
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 24
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 21
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 22
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 24
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 31
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 32
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 16
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 29
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 30
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 27
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 28
; SI-NEXT:    s_lshr_b64 vcc, s[6:7], 24
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 37
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 38
; SI-NEXT:    s_lshr_b64 vcc, s[6:7], 16
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 35
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 36
; SI-NEXT:    s_lshr_b64 vcc, s[6:7], 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 33
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 34
; SI-NEXT:    s_mov_b32 vcc_lo, s99
; SI-NEXT:    s_lshr_b64 s[98:99], s[4:5], 24
; SI-NEXT:    s_mov_b32 s99, s17
; SI-NEXT:    s_lshr_b64 s[16:17], s[4:5], 16
; SI-NEXT:    s_mov_b32 s17, s99
; SI-NEXT:    s_mov_b32 s99, vcc_lo
; SI-NEXT:    s_lshr_b64 vcc, s[4:5], 8
; SI-NEXT:    v_writelane_b32 v36, vcc_lo, 39
; SI-NEXT:    s_bfe_u32 s97, s36, 0x80008
; SI-NEXT:    s_bfe_u32 s85, s48, 0x80008
; SI-NEXT:    s_bfe_u32 s38, s30, 0x80008
; SI-NEXT:    s_bfe_u32 s86, s55, 0x80008
; SI-NEXT:    s_bfe_u32 s87, s67, 0x80008
; SI-NEXT:    s_bfe_u32 s50, s71, 0x80008
; SI-NEXT:    s_bfe_u32 s96, s82, 0x80008
; SI-NEXT:    v_writelane_b32 v36, vcc_hi, 40
; SI-NEXT:    s_cbranch_execnz .LBB95_4
; SI-NEXT:  .LBB95_2: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s83
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s84
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s82
; SI-NEXT:    v_readlane_b32 s12, v37, 12
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v3
; SI-NEXT:    v_readfirstlane_b32 s5, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s81
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s80
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_readfirstlane_b32 s4, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s6, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s34
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readfirstlane_b32 s5, v1
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_readfirstlane_b32 s6, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s71
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v2
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s70
; SI-NEXT:    v_readfirstlane_b32 s7, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s69
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readfirstlane_b32 s7, v2
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    v_readfirstlane_b32 s8, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s68
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_readfirstlane_b32 s8, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s67
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v3
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s66
; SI-NEXT:    v_readfirstlane_b32 s9, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s65
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    v_readfirstlane_b32 s10, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s64
; SI-NEXT:    v_readfirstlane_b32 s9, v3
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_readfirstlane_b32 s10, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s55
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v4
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s54
; SI-NEXT:    v_readfirstlane_b32 s11, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s12
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    v_readfirstlane_b32 s11, v4
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    v_readfirstlane_b32 s12, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s52
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_readfirstlane_b32 s12, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s30
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v5
; SI-NEXT:    v_readlane_b32 s14, v37, 11
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s51
; SI-NEXT:    v_readfirstlane_b32 s13, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s14
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    v_readfirstlane_b32 s14, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v6, s49
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    v_readfirstlane_b32 s13, v5
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_readfirstlane_b32 s14, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s48
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v6
; SI-NEXT:    v_readlane_b32 s16, v37, 10
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s39
; SI-NEXT:    v_readfirstlane_b32 s15, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s16
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    v_readfirstlane_b32 s15, v6
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    v_readfirstlane_b32 s40, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s37
; SI-NEXT:    s_or_b32 s15, s40, s15
; SI-NEXT:    v_readfirstlane_b32 s40, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s36
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v7
; SI-NEXT:    v_readlane_b32 s16, v37, 9
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s35
; SI-NEXT:    v_readfirstlane_b32 s41, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s16
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    s_lshl_b32 s40, s40, 16
; SI-NEXT:    v_readfirstlane_b32 s42, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v8, s31
; SI-NEXT:    s_or_b32 s40, s41, s40
; SI-NEXT:    v_readfirstlane_b32 s41, v7
; SI-NEXT:    s_lshl_b32 s41, s41, 16
; SI-NEXT:    s_or_b32 s41, s42, s41
; SI-NEXT:    v_readfirstlane_b32 s42, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s94
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v8
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s95
; SI-NEXT:    v_readfirstlane_b32 s43, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s99
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    s_or_b32 s42, s43, s42
; SI-NEXT:    v_readfirstlane_b32 s43, v8
; SI-NEXT:    s_lshl_b32 s43, s43, 16
; SI-NEXT:    v_readfirstlane_b32 s44, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v9, s93
; SI-NEXT:    s_or_b32 s43, s44, s43
; SI-NEXT:    v_readfirstlane_b32 s44, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s92
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v9
; SI-NEXT:    v_readlane_b32 s46, v37, 8
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s91
; SI-NEXT:    v_readfirstlane_b32 s45, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s46
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    v_readfirstlane_b32 s46, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v10, s28
; SI-NEXT:    v_readfirstlane_b32 s28, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s90
; SI-NEXT:    s_or_b32 s44, s45, s44
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v10
; SI-NEXT:    v_readfirstlane_b32 s45, v9
; SI-NEXT:    s_lshl_b32 s45, s45, 16
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s29
; SI-NEXT:    v_readlane_b32 s29, v37, 7
; SI-NEXT:    s_or_b32 s45, s46, s45
; SI-NEXT:    v_readfirstlane_b32 s46, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s29
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_readfirstlane_b32 s29, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v11, s26
; SI-NEXT:    s_lshl_b32 s28, s28, 16
; SI-NEXT:    v_readfirstlane_b32 s26, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s89
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v11
; SI-NEXT:    s_or_b32 s46, s46, s28
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s27
; SI-NEXT:    v_readfirstlane_b32 s28, v10
; SI-NEXT:    s_lshl_b32 s28, s28, 16
; SI-NEXT:    v_readlane_b32 s27, v37, 6
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    s_or_b32 s47, s29, s28
; SI-NEXT:    v_readfirstlane_b32 s28, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s27
; SI-NEXT:    v_readfirstlane_b32 s27, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s24
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    s_lshl_b32 s26, s26, 16
; SI-NEXT:    s_or_b32 s60, s28, s26
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_readfirstlane_b32 s26, v11
; SI-NEXT:    s_lshl_b32 s26, s26, 16
; SI-NEXT:    s_or_b32 s61, s27, s26
; SI-NEXT:    v_readfirstlane_b32 s26, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s25
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_readlane_b32 s25, v37, 5
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_readfirstlane_b32 s24, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v13, s88
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s25
; SI-NEXT:    v_readfirstlane_b32 s25, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s22
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    s_lshl_b32 s24, s24, 16
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    s_or_b32 s56, s26, s24
; SI-NEXT:    v_readfirstlane_b32 s24, v13
; SI-NEXT:    s_lshl_b32 s24, s24, 16
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    s_or_b32 s57, s25, s24
; SI-NEXT:    v_readfirstlane_b32 s24, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s23
; SI-NEXT:    v_readfirstlane_b32 s22, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s79
; SI-NEXT:    v_readlane_b32 s23, v37, 4
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s23
; SI-NEXT:    v_readfirstlane_b32 s23, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s20
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    s_or_b32 s58, s24, s22
; SI-NEXT:    v_readfirstlane_b32 s22, v16
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_or_b32 s59, s23, s22
; SI-NEXT:    v_readfirstlane_b32 s20, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s78
; SI-NEXT:    v_readfirstlane_b32 s22, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s21
; SI-NEXT:    v_readlane_b32 s21, v37, 2
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v14
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s21
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    s_lshr_b64 s[64:65], s[14:15], 24
; SI-NEXT:    v_writelane_b32 v36, s64, 13
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_readfirstlane_b32 s21, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s18
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_writelane_b32 v36, s65, 14
; SI-NEXT:    s_lshr_b64 s[64:65], s[12:13], 24
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_readfirstlane_b32 s18, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s77
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    v_writelane_b32 v36, s64, 19
; SI-NEXT:    s_or_b32 s62, s22, s20
; SI-NEXT:    v_readfirstlane_b32 s20, v20
; SI-NEXT:    v_writelane_b32 v36, s65, 20
; SI-NEXT:    s_lshr_b64 s[64:65], s[12:13], 16
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    v_writelane_b32 v36, s64, 17
; SI-NEXT:    s_or_b32 s63, s21, s20
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_readfirstlane_b32 s20, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v12, s19
; SI-NEXT:    v_readlane_b32 s19, v37, 0
; SI-NEXT:    v_writelane_b32 v36, s65, 18
; SI-NEXT:    s_lshr_b64 s[64:65], s[12:13], 8
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s19
; SI-NEXT:    v_writelane_b32 v36, s64, 15
; SI-NEXT:    v_writelane_b32 v36, s65, 16
; SI-NEXT:    s_lshr_b64 s[84:85], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v36, s84, 23
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_writelane_b32 v36, s85, 24
; SI-NEXT:    s_lshr_b64 s[84:85], s[10:11], 8
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s53
; SI-NEXT:    v_writelane_b32 v36, s84, 21
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_writelane_b32 v36, s85, 22
; SI-NEXT:    s_lshr_b64 s[84:85], s[8:9], 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    v_writelane_b32 v36, s84, 31
; SI-NEXT:    s_or_b32 s72, s20, s18
; SI-NEXT:    v_readfirstlane_b32 s18, v23
; SI-NEXT:    v_writelane_b32 v36, s85, 32
; SI-NEXT:    s_lshr_b64 s[84:85], s[8:9], 16
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    v_readfirstlane_b32 s16, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v15
; SI-NEXT:    v_writelane_b32 v36, s84, 29
; SI-NEXT:    s_or_b32 s73, s16, s18
; SI-NEXT:    v_readfirstlane_b32 s16, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v14, s76
; SI-NEXT:    v_writelane_b32 v36, s85, 30
; SI-NEXT:    s_lshr_b64 s[84:85], s[8:9], 8
; SI-NEXT:    v_cvt_f32_f16_e32 v15, s17
; SI-NEXT:    v_writelane_b32 v36, s84, 27
; SI-NEXT:    s_lshr_b64 s[20:21], s[46:47], 24
; SI-NEXT:    v_writelane_b32 v36, s85, 28
; SI-NEXT:    s_lshr_b64 s[84:85], s[6:7], 24
; SI-NEXT:    v_writelane_b32 v37, s20, 53
; SI-NEXT:    v_writelane_b32 v36, s84, 37
; SI-NEXT:    v_readfirstlane_b32 s17, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v14
; SI-NEXT:    v_writelane_b32 v37, s21, 54
; SI-NEXT:    s_lshr_b64 s[20:21], s[46:47], 16
; SI-NEXT:    v_writelane_b32 v36, s85, 38
; SI-NEXT:    s_lshr_b64 s[84:85], s[6:7], 16
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v15
; SI-NEXT:    v_writelane_b32 v37, s20, 51
; SI-NEXT:    v_writelane_b32 v36, s84, 35
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_writelane_b32 v37, s21, 52
; SI-NEXT:    s_lshr_b64 s[20:21], s[46:47], 8
; SI-NEXT:    v_writelane_b32 v36, s85, 36
; SI-NEXT:    s_lshr_b64 s[84:85], s[6:7], 8
; SI-NEXT:    v_writelane_b32 v37, s20, 49
; SI-NEXT:    v_writelane_b32 v36, s84, 33
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    v_writelane_b32 v37, s21, 50
; SI-NEXT:    v_writelane_b32 v36, s85, 34
; SI-NEXT:    s_lshr_b64 s[84:85], s[4:5], 24
; SI-NEXT:    s_or_b32 s74, s17, s16
; SI-NEXT:    v_readfirstlane_b32 s16, v26
; SI-NEXT:    v_writelane_b32 v37, s84, 0
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    v_readfirstlane_b32 s17, v12
; SI-NEXT:    v_writelane_b32 v37, s85, 1
; SI-NEXT:    s_lshr_b64 s[84:85], s[4:5], 16
; SI-NEXT:    s_or_b32 s75, s17, s16
; SI-NEXT:    v_writelane_b32 v37, s84, 2
; SI-NEXT:    s_lshr_b64 s[16:17], s[74:75], 8
; SI-NEXT:    v_writelane_b32 v37, s85, 3
; SI-NEXT:    s_lshr_b64 s[84:85], s[4:5], 8
; SI-NEXT:    s_lshr_b64 s[18:19], s[74:75], 24
; SI-NEXT:    s_lshr_b64 s[38:39], s[72:73], 16
; SI-NEXT:    s_lshr_b64 s[52:53], s[72:73], 8
; SI-NEXT:    s_lshr_b64 s[88:89], s[62:63], 24
; SI-NEXT:    s_lshr_b64 s[94:95], s[62:63], 16
; SI-NEXT:    s_lshr_b64 s[36:37], s[62:63], 8
; SI-NEXT:    s_lshr_b64 s[76:77], s[58:59], 24
; SI-NEXT:    s_lshr_b64 s[78:79], s[58:59], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[58:59], 8
; SI-NEXT:    s_lshr_b64 s[34:35], s[60:61], 24
; SI-NEXT:    s_lshr_b64 s[50:51], s[60:61], 16
; SI-NEXT:    s_lshr_b64 s[22:23], s[60:61], 8
; SI-NEXT:    s_lshr_b32 s20, s45, 8
; SI-NEXT:    s_lshr_b32 s17, s15, 8
; SI-NEXT:    s_lshr_b64 s[26:27], s[40:41], 24
; SI-NEXT:    s_lshr_b64 s[82:83], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[64:65], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v36, s84, 39
; SI-NEXT:    s_lshr_b64 s[54:55], s[74:75], 16
; SI-NEXT:    s_lshr_b64 vcc, s[72:73], 24
; SI-NEXT:    s_lshr_b64 s[70:71], s[56:57], 24
; SI-NEXT:    s_lshr_b64 s[86:87], s[56:57], 16
; SI-NEXT:    s_lshr_b64 s[98:99], s[56:57], 8
; SI-NEXT:    s_lshr_b32 s23, s75, 8
; SI-NEXT:    s_lshr_b32 s91, s73, 8
; SI-NEXT:    s_lshr_b32 s89, s63, 8
; SI-NEXT:    s_lshr_b32 s79, s59, 8
; SI-NEXT:    s_lshr_b32 s77, s57, 8
; SI-NEXT:    s_lshr_b32 s29, s61, 8
; SI-NEXT:    s_lshr_b32 s28, s47, 8
; SI-NEXT:    s_lshr_b32 s19, s43, 8
; SI-NEXT:    s_lshr_b32 s53, s41, 8
; SI-NEXT:    s_lshr_b32 s51, s13, 8
; SI-NEXT:    s_lshr_b32 s39, s11, 8
; SI-NEXT:    s_lshr_b32 s37, s9, 8
; SI-NEXT:    s_lshr_b32 s35, s7, 8
; SI-NEXT:    s_lshr_b32 s95, s5, 8
; SI-NEXT:    v_bfe_u32 v32, v26, 8, 8
; SI-NEXT:    v_bfe_u32 v31, v23, 8, 8
; SI-NEXT:    v_bfe_u32 v30, v20, 8, 8
; SI-NEXT:    v_bfe_u32 v29, v16, 8, 8
; SI-NEXT:    v_bfe_u32 v28, v13, 8, 8
; SI-NEXT:    v_bfe_u32 v27, v11, 8, 8
; SI-NEXT:    v_bfe_u32 v25, v10, 8, 8
; SI-NEXT:    v_bfe_u32 v24, v9, 8, 8
; SI-NEXT:    v_bfe_u32 v22, v8, 8, 8
; SI-NEXT:    v_bfe_u32 v21, v7, 8, 8
; SI-NEXT:    v_bfe_u32 v19, v6, 8, 8
; SI-NEXT:    v_bfe_u32 v18, v5, 8, 8
; SI-NEXT:    v_bfe_u32 v17, v4, 8, 8
; SI-NEXT:    v_bfe_u32 v15, v3, 8, 8
; SI-NEXT:    v_bfe_u32 v14, v2, 8, 8
; SI-NEXT:    v_bfe_u32 v12, v1, 8, 8
; SI-NEXT:    s_lshr_b64 s[68:69], s[44:45], 24
; SI-NEXT:    s_lshr_b64 s[80:81], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[48:49], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[92:93], s[42:43], 24
; SI-NEXT:    s_lshr_b64 s[24:25], s[42:43], 16
; SI-NEXT:    s_lshr_b64 s[30:31], s[42:43], 8
; SI-NEXT:    s_mov_b32 s27, s20
; SI-NEXT:    s_lshr_b64 s[96:97], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[20:21], s[40:41], 8
; SI-NEXT:    s_lshr_b64 s[66:67], s[14:15], 8
; SI-NEXT:    v_writelane_b32 v36, s85, 40
; SI-NEXT:    s_mov_b32 s84, s64
; SI-NEXT:    s_mov_b32 s64, s82
; SI-NEXT:    s_mov_b32 s82, s26
; SI-NEXT:    s_mov_b32 s26, s17
; SI-NEXT:    s_branch .LBB95_5
; SI-NEXT:  .LBB95_3:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr16
; SI-NEXT:    v_writelane_b32 v37, s4, 13
; SI-NEXT:    v_writelane_b32 v37, s5, 14
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr97
; SI-NEXT:    ; implicit-def: $sgpr85
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr87
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    v_writelane_b32 v37, s4, 15
; SI-NEXT:    v_writelane_b32 v37, s5, 16
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 17
; SI-NEXT:    v_writelane_b32 v37, s5, 18
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 19
; SI-NEXT:    v_writelane_b32 v37, s5, 20
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 21
; SI-NEXT:    v_writelane_b32 v37, s5, 22
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 23
; SI-NEXT:    v_writelane_b32 v37, s5, 24
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 25
; SI-NEXT:    v_writelane_b32 v37, s5, 26
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 27
; SI-NEXT:    v_writelane_b32 v37, s5, 28
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 29
; SI-NEXT:    v_writelane_b32 v37, s5, 30
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 31
; SI-NEXT:    v_writelane_b32 v37, s5, 32
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 33
; SI-NEXT:    v_writelane_b32 v37, s5, 34
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 35
; SI-NEXT:    v_writelane_b32 v37, s5, 36
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 37
; SI-NEXT:    v_writelane_b32 v37, s5, 38
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 39
; SI-NEXT:    v_writelane_b32 v37, s5, 40
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 41
; SI-NEXT:    v_writelane_b32 v37, s5, 42
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 43
; SI-NEXT:    v_writelane_b32 v37, s5, 44
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 45
; SI-NEXT:    v_writelane_b32 v37, s5, 46
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 47
; SI-NEXT:    v_writelane_b32 v37, s5, 48
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 49
; SI-NEXT:    v_writelane_b32 v37, s5, 50
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 51
; SI-NEXT:    v_writelane_b32 v37, s5, 52
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 53
; SI-NEXT:    v_writelane_b32 v37, s5, 54
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 55
; SI-NEXT:    v_writelane_b32 v37, s5, 56
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 57
; SI-NEXT:    v_writelane_b32 v37, s5, 58
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 59
; SI-NEXT:    v_writelane_b32 v37, s5, 60
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 61
; SI-NEXT:    v_writelane_b32 v37, s5, 62
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v37, s4, 63
; SI-NEXT:    v_writelane_b32 v36, s5, 0
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 1
; SI-NEXT:    v_writelane_b32 v36, s5, 2
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 3
; SI-NEXT:    v_writelane_b32 v36, s5, 4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 5
; SI-NEXT:    v_writelane_b32 v36, s5, 6
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 7
; SI-NEXT:    v_writelane_b32 v36, s5, 8
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 9
; SI-NEXT:    v_writelane_b32 v36, s5, 10
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 11
; SI-NEXT:    v_writelane_b32 v36, s5, 12
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 13
; SI-NEXT:    v_writelane_b32 v36, s5, 14
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 15
; SI-NEXT:    v_writelane_b32 v36, s5, 16
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 17
; SI-NEXT:    v_writelane_b32 v36, s5, 18
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 19
; SI-NEXT:    v_writelane_b32 v36, s5, 20
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 21
; SI-NEXT:    v_writelane_b32 v36, s5, 22
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 23
; SI-NEXT:    v_writelane_b32 v36, s5, 24
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 25
; SI-NEXT:    v_writelane_b32 v36, s5, 26
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 27
; SI-NEXT:    v_writelane_b32 v36, s5, 28
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 29
; SI-NEXT:    v_writelane_b32 v36, s5, 30
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 31
; SI-NEXT:    v_writelane_b32 v36, s5, 32
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 33
; SI-NEXT:    v_writelane_b32 v36, s5, 34
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 35
; SI-NEXT:    v_writelane_b32 v36, s5, 36
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s4, 37
; SI-NEXT:    v_writelane_b32 v36, s5, 38
; SI-NEXT:    v_writelane_b32 v36, s16, 39
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v36, s17, 40
; SI-NEXT:    ; implicit-def: $sgpr16
; SI-NEXT:    s_branch .LBB95_2
; SI-NEXT:  .LBB95_4:
; SI-NEXT:    v_readlane_b32 s17, v36, 41
; SI-NEXT:    v_mov_b32_e32 v22, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 42
; SI-NEXT:    v_mov_b32_e32 v24, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 43
; SI-NEXT:    v_mov_b32_e32 v25, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 44
; SI-NEXT:    v_mov_b32_e32 v27, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 45
; SI-NEXT:    v_mov_b32_e32 v10, s90
; SI-NEXT:    v_mov_b32_e32 v11, s89
; SI-NEXT:    v_mov_b32_e32 v13, s88
; SI-NEXT:    v_mov_b32_e32 v16, s79
; SI-NEXT:    v_mov_b32_e32 v20, s78
; SI-NEXT:    v_mov_b32_e32 v23, s77
; SI-NEXT:    v_mov_b32_e32 v26, s76
; SI-NEXT:    v_mov_b32_e32 v28, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 46
; SI-NEXT:    v_readlane_b32 s76, v37, 35
; SI-NEXT:    v_readlane_b32 s78, v37, 33
; SI-NEXT:    v_readlane_b32 s88, v37, 29
; SI-NEXT:    v_readlane_b32 s90, v37, 31
; SI-NEXT:    v_mov_b32_e32 v29, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 47
; SI-NEXT:    v_readlane_b32 s77, v37, 36
; SI-NEXT:    v_readlane_b32 s79, v37, 34
; SI-NEXT:    v_readlane_b32 s89, v37, 30
; SI-NEXT:    v_readlane_b32 s91, v37, 32
; SI-NEXT:    v_writelane_b32 v37, s98, 0
; SI-NEXT:    v_mov_b32_e32 v30, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 48
; SI-NEXT:    v_writelane_b32 v37, s99, 1
; SI-NEXT:    v_mov_b32_e32 v8, s94
; SI-NEXT:    v_mov_b32_e32 v31, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 49
; SI-NEXT:    v_readlane_b32 s94, v37, 27
; SI-NEXT:    v_mov_b32_e32 v32, s17
; SI-NEXT:    v_readlane_b32 s17, v36, 50
; SI-NEXT:    v_readlane_b32 s95, v37, 28
; SI-NEXT:    s_mov_b32 s95, s17
; SI-NEXT:    v_writelane_b32 v37, s16, 2
; SI-NEXT:    v_writelane_b32 v37, s17, 3
; SI-NEXT:    v_mov_b32_e32 v14, s50
; SI-NEXT:    v_readlane_b32 s34, v37, 47
; SI-NEXT:    v_readlane_b32 s50, v37, 45
; SI-NEXT:    v_mov_b32_e32 v7, s36
; SI-NEXT:    v_mov_b32_e32 v15, s87
; SI-NEXT:    v_mov_b32_e32 v17, s86
; SI-NEXT:    v_mov_b32_e32 v18, s38
; SI-NEXT:    v_readlane_b32 s22, v36, 51
; SI-NEXT:    v_readlane_b32 s23, v36, 54
; SI-NEXT:    v_readlane_b32 s36, v37, 25
; SI-NEXT:    v_readlane_b32 s38, v37, 21
; SI-NEXT:    v_readlane_b32 s16, v37, 17
; SI-NEXT:    v_readlane_b32 s52, v37, 19
; SI-NEXT:    v_readlane_b32 s86, v36, 25
; SI-NEXT:    v_readlane_b32 s35, v37, 48
; SI-NEXT:    v_readlane_b32 s51, v37, 46
; SI-NEXT:    v_readlane_b32 s18, v36, 52
; SI-NEXT:    v_readlane_b32 s20, v36, 53
; SI-NEXT:    v_readlane_b32 s21, v36, 56
; SI-NEXT:    v_readlane_b32 s37, v37, 26
; SI-NEXT:    v_readlane_b32 s39, v37, 22
; SI-NEXT:    v_readlane_b32 s17, v37, 18
; SI-NEXT:    v_readlane_b32 s53, v37, 20
; SI-NEXT:    v_readlane_b32 s87, v36, 26
; SI-NEXT:    s_mov_b32 s35, s22
; SI-NEXT:    s_mov_b32 s51, s23
; SI-NEXT:    v_readlane_b32 s22, v37, 43
; SI-NEXT:    v_mov_b32_e32 v1, s82
; SI-NEXT:    v_mov_b32_e32 v2, s71
; SI-NEXT:    v_mov_b32_e32 v3, s67
; SI-NEXT:    v_mov_b32_e32 v4, s55
; SI-NEXT:    v_mov_b32_e32 v5, s30
; SI-NEXT:    v_mov_b32_e32 v6, s48
; SI-NEXT:    v_mov_b32_e32 v9, s92
; SI-NEXT:    v_mov_b32_e32 v12, s96
; SI-NEXT:    v_mov_b32_e32 v21, s97
; SI-NEXT:    v_readlane_b32 vcc_lo, v37, 23
; SI-NEXT:    s_mov_b32 s37, s18
; SI-NEXT:    s_mov_b32 s39, s20
; SI-NEXT:    s_mov_b32 s18, s16
; SI-NEXT:    s_mov_b32 s53, s21
; SI-NEXT:    v_readlane_b32 s54, v37, 15
; SI-NEXT:    v_readlane_b32 s16, v37, 13
; SI-NEXT:    s_mov_b32 s84, s86
; SI-NEXT:    v_readlane_b32 s64, v36, 11
; SI-NEXT:    v_readlane_b32 s23, v37, 44
; SI-NEXT:    v_readlane_b32 s66, v36, 9
; SI-NEXT:    v_readlane_b32 s70, v37, 41
; SI-NEXT:    v_readlane_b32 s82, v36, 7
; SI-NEXT:    v_readlane_b32 s86, v37, 39
; SI-NEXT:    v_readlane_b32 s96, v36, 5
; SI-NEXT:    v_readlane_b32 s98, v37, 37
; SI-NEXT:    v_readlane_b32 s20, v36, 3
; SI-NEXT:    v_readlane_b32 s92, v36, 1
; SI-NEXT:    v_readlane_b32 s24, v37, 63
; SI-NEXT:    v_readlane_b32 s30, v37, 61
; SI-NEXT:    v_readlane_b32 s68, v37, 59
; SI-NEXT:    v_readlane_b32 s80, v37, 57
; SI-NEXT:    v_readlane_b32 s48, v37, 55
; SI-NEXT:    v_mov_b32_e32 v19, s85
; SI-NEXT:    v_readlane_b32 s26, v36, 55
; SI-NEXT:    v_readlane_b32 s19, v36, 57
; SI-NEXT:    v_readlane_b32 s27, v36, 58
; SI-NEXT:    v_readlane_b32 s28, v36, 59
; SI-NEXT:    v_readlane_b32 s29, v36, 60
; SI-NEXT:    v_readlane_b32 vcc_hi, v37, 24
; SI-NEXT:    v_readlane_b32 s77, v36, 61
; SI-NEXT:    v_readlane_b32 s55, v37, 16
; SI-NEXT:    v_readlane_b32 s17, v37, 14
; SI-NEXT:    v_readlane_b32 s79, v36, 62
; SI-NEXT:    v_readlane_b32 s89, v36, 63
; SI-NEXT:    v_readlane_b32 s91, v35, 0
; SI-NEXT:    v_readlane_b32 s65, v36, 12
; SI-NEXT:    v_readlane_b32 s67, v36, 10
; SI-NEXT:    v_readlane_b32 s23, v35, 1
; SI-NEXT:    v_readlane_b32 s71, v37, 42
; SI-NEXT:    v_readlane_b32 s83, v36, 8
; SI-NEXT:    v_readlane_b32 s87, v37, 40
; SI-NEXT:    v_readlane_b32 s97, v36, 6
; SI-NEXT:    v_readlane_b32 s99, v37, 38
; SI-NEXT:    v_readlane_b32 s21, v36, 4
; SI-NEXT:    v_readlane_b32 s93, v36, 2
; SI-NEXT:    v_readlane_b32 s25, v36, 0
; SI-NEXT:    v_readlane_b32 s31, v37, 62
; SI-NEXT:    v_readlane_b32 s69, v37, 60
; SI-NEXT:    v_readlane_b32 s81, v37, 58
; SI-NEXT:    v_readlane_b32 s49, v37, 56
; SI-NEXT:  .LBB95_5: ; %end
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s74, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s54, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v33, s16
; SI-NEXT:    s_and_b32 s16, s75, 0xff
; SI-NEXT:    s_lshl_b32 s17, s23, 8
; SI-NEXT:    v_and_b32_e32 v26, 0xff, v26
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_lshlrev_b32_e32 v32, 24, v32
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v26, v32, v26
; SI-NEXT:    v_or_b32_e32 v26, s16, v26
; SI-NEXT:    s_lshl_b32 s16, s52, 8
; SI-NEXT:    s_and_b32 s17, s72, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s38, 0xff
; SI-NEXT:    s_lshl_b32 s18, vcc_lo, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v32, s16
; SI-NEXT:    s_and_b32 s16, s73, 0xff
; SI-NEXT:    s_lshl_b32 s17, s91, 8
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v23
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_lshlrev_b32_e32 v31, 24, v31
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v23, v31, v23
; SI-NEXT:    v_or_b32_e32 v23, s16, v23
; SI-NEXT:    s_lshl_b32 s16, s36, 8
; SI-NEXT:    s_and_b32 s17, s62, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s94, 0xff
; SI-NEXT:    s_lshl_b32 s18, s88, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v31, s16
; SI-NEXT:    s_and_b32 s16, s63, 0xff
; SI-NEXT:    s_lshl_b32 s17, s89, 8
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_lshlrev_b32_e32 v30, 24, v30
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v20, v30, v20
; SI-NEXT:    v_or_b32_e32 v20, s16, v20
; SI-NEXT:    s_lshl_b32 s16, s90, 8
; SI-NEXT:    s_and_b32 s17, s58, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s78, 0xff
; SI-NEXT:    s_lshl_b32 s18, s76, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v30, s16
; SI-NEXT:    s_and_b32 s16, s59, 0xff
; SI-NEXT:    s_lshl_b32 s17, s79, 8
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v29, 24, v29
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v16, v29, v16
; SI-NEXT:    v_or_b32_e32 v16, s16, v16
; SI-NEXT:    s_lshl_b32 s16, s98, 8
; SI-NEXT:    s_and_b32 s17, s56, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s86, 0xff
; SI-NEXT:    s_lshl_b32 s18, s70, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v29, s16
; SI-NEXT:    s_and_b32 s16, s57, 0xff
; SI-NEXT:    s_lshl_b32 s17, s77, 8
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_lshlrev_b32_e32 v28, 24, v28
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v13, v28, v13
; SI-NEXT:    v_or_b32_e32 v13, s16, v13
; SI-NEXT:    s_lshl_b32 s16, s22, 8
; SI-NEXT:    s_and_b32 s17, s60, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s50, 0xff
; SI-NEXT:    s_lshl_b32 s18, s34, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v27, 24, v27
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_or_b32_e32 v11, v27, v11
; SI-NEXT:    v_add_i32_e32 v27, vcc, 4, v0
; SI-NEXT:    v_mov_b32_e32 v28, s16
; SI-NEXT:    s_and_b32 s16, s61, 0xff
; SI-NEXT:    s_lshl_b32 s17, s29, 8
; SI-NEXT:    buffer_store_dword v33, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_store_dword v26, v27, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v26, vcc, 8, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v32, v26, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v26, vcc, 12, v0
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    buffer_store_dword v23, v26, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v23, vcc, 16, v0
; SI-NEXT:    v_or_b32_e32 v11, s16, v11
; SI-NEXT:    buffer_store_dword v31, v23, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v23, vcc, 20, v0
; SI-NEXT:    v_readlane_b32 s16, v37, 49
; SI-NEXT:    buffer_store_dword v20, v23, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 24, v0
; SI-NEXT:    v_readlane_b32 s17, v37, 50
; SI-NEXT:    v_readlane_b32 s22, v37, 51
; SI-NEXT:    buffer_store_dword v30, v20, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v20, vcc, 28, v0
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s46, 0xff
; SI-NEXT:    v_readlane_b32 s23, v37, 52
; SI-NEXT:    buffer_store_dword v16, v20, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 32, v0
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s22, 0xff
; SI-NEXT:    v_readlane_b32 s22, v37, 53
; SI-NEXT:    buffer_store_dword v29, v16, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v16, vcc, 36, v0
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s22, 24
; SI-NEXT:    buffer_store_dword v13, v16, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v13, vcc, 40, v0
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    buffer_store_dword v28, v13, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v13, vcc, 44, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v11, v13, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v11, vcc, 48, v0
; SI-NEXT:    v_mov_b32_e32 v13, s16
; SI-NEXT:    s_and_b32 s16, s47, 0xff
; SI-NEXT:    s_lshl_b32 s17, s28, 8
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    buffer_store_dword v13, v11, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshlrev_b32_e32 v11, 24, v25
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    v_or_b32_e32 v10, s16, v10
; SI-NEXT:    s_and_b32 s16, s44, 0xff
; SI-NEXT:    s_lshl_b32 s17, s48, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s80, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s68, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v11, vcc, 52, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v10, v11, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v10, vcc, 56, v0
; SI-NEXT:    v_mov_b32_e32 v11, s16
; SI-NEXT:    s_and_b32 s16, s45, 0xff
; SI-NEXT:    s_lshl_b32 s17, s27, 8
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    buffer_store_dword v11, v10, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_lshlrev_b32_e32 v10, 24, v24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_or_b32_e32 v9, s16, v9
; SI-NEXT:    s_and_b32 s16, s42, 0xff
; SI-NEXT:    s_lshl_b32 s17, s30, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s24, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s92, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v10, vcc, 60, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v9, v10, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v9, vcc, 64, v0
; SI-NEXT:    v_mov_b32_e32 v10, s16
; SI-NEXT:    s_and_b32 s16, s43, 0xff
; SI-NEXT:    s_lshl_b32 s17, s19, 8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    buffer_store_dword v10, v9, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_lshlrev_b32_e32 v9, 24, v22
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_or_b32_e32 v8, s16, v8
; SI-NEXT:    s_and_b32 s16, s40, 0xff
; SI-NEXT:    s_lshl_b32 s17, s20, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s96, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s82, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v9, vcc, 0x44, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v8, v9, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 0x48, v0
; SI-NEXT:    v_mov_b32_e32 v9, s16
; SI-NEXT:    s_and_b32 s16, s41, 0xff
; SI-NEXT:    s_lshl_b32 s17, s53, 8
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    buffer_store_dword v9, v8, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_lshlrev_b32_e32 v8, 24, v21
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_or_b32_e32 v7, s16, v7
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s16, s66, 8
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    s_and_b32 s16, s64, 0xff
; SI-NEXT:    v_readlane_b32 s20, v36, 13
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s17, s20, 24
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_add_i32_e32 v8, vcc, 0x4c, v0
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    buffer_store_dword v7, v8, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 0x50, v0
; SI-NEXT:    v_mov_b32_e32 v8, s14
; SI-NEXT:    s_and_b32 s14, s15, 0xff
; SI-NEXT:    s_lshl_b32 s15, s26, 8
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    buffer_store_dword v8, v7, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_lshlrev_b32_e32 v7, 24, v19
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    v_or_b32_e32 v6, s14, v6
; SI-NEXT:    v_readlane_b32 s14, v36, 15
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s15, v36, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 8
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    v_readlane_b32 s14, v36, 17
; SI-NEXT:    v_readlane_b32 s15, v36, 18
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s16, v36, 19
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_lshl_b32 s15, s16, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    v_add_i32_e32 v7, vcc, 0x54, v0
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    buffer_store_dword v6, v7, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x58, v0
; SI-NEXT:    v_mov_b32_e32 v7, s12
; SI-NEXT:    s_and_b32 s12, s13, 0xff
; SI-NEXT:    s_lshl_b32 s13, s51, 8
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    buffer_store_dword v7, v6, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_lshlrev_b32_e32 v6, 24, v18
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_or_b32_e32 v5, s12, v5
; SI-NEXT:    v_readlane_b32 s12, v36, 21
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s13, v36, 22
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    v_readlane_b32 s12, v36, 23
; SI-NEXT:    v_readlane_b32 s13, v36, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s13, s84, 24
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    v_add_i32_e32 v6, vcc, 0x5c, v0
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    buffer_store_dword v5, v6, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 0x60, v0
; SI-NEXT:    v_mov_b32_e32 v6, s10
; SI-NEXT:    s_and_b32 s10, s11, 0xff
; SI-NEXT:    s_lshl_b32 s11, s39, 8
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    buffer_store_dword v6, v5, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v17
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    v_or_b32_e32 v4, s10, v4
; SI-NEXT:    v_readlane_b32 s10, v36, 27
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s11, v36, 28
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    v_readlane_b32 s10, v36, 29
; SI-NEXT:    v_readlane_b32 s11, v36, 30
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s12, v36, 31
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s11, s12, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    v_add_i32_e32 v5, vcc, 0x64, v0
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    buffer_store_dword v4, v5, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x68, v0
; SI-NEXT:    v_mov_b32_e32 v5, s8
; SI-NEXT:    s_and_b32 s8, s9, 0xff
; SI-NEXT:    s_lshl_b32 s9, s37, 8
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    buffer_store_dword v5, v4, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshlrev_b32_e32 v4, 24, v15
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_or_b32_e32 v3, s8, v3
; SI-NEXT:    v_readlane_b32 s8, v36, 33
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s9, v36, 34
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    v_readlane_b32 s8, v36, 35
; SI-NEXT:    v_readlane_b32 s9, v36, 36
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v36, 37
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s10, 24
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    v_add_i32_e32 v4, vcc, 0x6c, v0
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    buffer_store_dword v3, v4, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x70, v0
; SI-NEXT:    v_mov_b32_e32 v4, s6
; SI-NEXT:    s_and_b32 s6, s7, 0xff
; SI-NEXT:    s_lshl_b32 s7, s35, 8
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    buffer_store_dword v4, v3, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v14
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v2, s6, v2
; SI-NEXT:    v_readlane_b32 s6, v36, 39
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    v_readlane_b32 s7, v36, 40
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    v_readlane_b32 s6, v37, 2
; SI-NEXT:    v_readlane_b32 s7, v37, 3
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s8, v37, 0
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s8, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_add_i32_e32 v3, vcc, 0x74, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    v_mov_b32_e32 v3, s4
; SI-NEXT:    s_and_b32 s4, s5, 0xff
; SI-NEXT:    s_lshl_b32 s5, s95, 8
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    buffer_store_dword v3, v2, s[0:3], 0 offen
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v12
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v1, s4, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    v_readlane_b32 s23, v37, 54
; SI-NEXT:    v_readlane_b32 s21, v36, 14
; SI-NEXT:    v_readlane_b32 s17, v36, 20
; SI-NEXT:    v_readlane_b32 s13, v36, 32
; SI-NEXT:    v_readlane_b32 s11, v36, 38
; SI-NEXT:    v_readlane_b32 s9, v37, 1
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    v_readlane_b32 s99, v34, 35
; SI-NEXT:    v_readlane_b32 s98, v34, 34
; SI-NEXT:    v_readlane_b32 s97, v34, 33
; SI-NEXT:    v_readlane_b32 s96, v34, 32
; SI-NEXT:    v_readlane_b32 s87, v34, 31
; SI-NEXT:    v_readlane_b32 s86, v34, 30
; SI-NEXT:    v_readlane_b32 s85, v34, 29
; SI-NEXT:    v_readlane_b32 s84, v34, 28
; SI-NEXT:    v_readlane_b32 s83, v34, 27
; SI-NEXT:    v_readlane_b32 s82, v34, 26
; SI-NEXT:    v_readlane_b32 s81, v34, 25
; SI-NEXT:    v_readlane_b32 s80, v34, 24
; SI-NEXT:    v_readlane_b32 s71, v34, 23
; SI-NEXT:    v_readlane_b32 s70, v34, 22
; SI-NEXT:    v_readlane_b32 s69, v34, 21
; SI-NEXT:    v_readlane_b32 s68, v34, 20
; SI-NEXT:    v_readlane_b32 s67, v34, 19
; SI-NEXT:    v_readlane_b32 s66, v34, 18
; SI-NEXT:    v_readlane_b32 s65, v34, 17
; SI-NEXT:    v_readlane_b32 s64, v34, 16
; SI-NEXT:    v_readlane_b32 s55, v34, 15
; SI-NEXT:    v_readlane_b32 s54, v34, 14
; SI-NEXT:    v_readlane_b32 s53, v34, 13
; SI-NEXT:    v_readlane_b32 s52, v34, 12
; SI-NEXT:    v_readlane_b32 s51, v34, 11
; SI-NEXT:    v_readlane_b32 s50, v34, 10
; SI-NEXT:    v_readlane_b32 s49, v34, 9
; SI-NEXT:    v_readlane_b32 s48, v34, 8
; SI-NEXT:    v_readlane_b32 s39, v34, 7
; SI-NEXT:    v_readlane_b32 s38, v34, 6
; SI-NEXT:    v_readlane_b32 s37, v34, 5
; SI-NEXT:    v_readlane_b32 s36, v34, 4
; SI-NEXT:    v_readlane_b32 s35, v34, 3
; SI-NEXT:    v_readlane_b32 s34, v34, 2
; SI-NEXT:    v_readlane_b32 s31, v34, 1
; SI-NEXT:    v_readlane_b32 s30, v34, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v128i8_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v63, s30, 0
; VI-NEXT:    v_writelane_b32 v63, s31, 1
; VI-NEXT:    v_writelane_b32 v63, s34, 2
; VI-NEXT:    v_writelane_b32 v63, s35, 3
; VI-NEXT:    v_writelane_b32 v63, s36, 4
; VI-NEXT:    v_writelane_b32 v63, s37, 5
; VI-NEXT:    v_writelane_b32 v63, s38, 6
; VI-NEXT:    v_writelane_b32 v63, s39, 7
; VI-NEXT:    v_writelane_b32 v63, s48, 8
; VI-NEXT:    v_writelane_b32 v63, s49, 9
; VI-NEXT:    v_writelane_b32 v63, s50, 10
; VI-NEXT:    v_writelane_b32 v63, s51, 11
; VI-NEXT:    v_writelane_b32 v63, s52, 12
; VI-NEXT:    v_writelane_b32 v63, s53, 13
; VI-NEXT:    v_writelane_b32 v63, s54, 14
; VI-NEXT:    v_writelane_b32 v63, s55, 15
; VI-NEXT:    v_writelane_b32 v63, s64, 16
; VI-NEXT:    v_writelane_b32 v63, s65, 17
; VI-NEXT:    v_writelane_b32 v63, s66, 18
; VI-NEXT:    v_writelane_b32 v63, s67, 19
; VI-NEXT:    v_writelane_b32 v63, s68, 20
; VI-NEXT:    v_writelane_b32 v63, s69, 21
; VI-NEXT:    v_writelane_b32 v63, s70, 22
; VI-NEXT:    v_writelane_b32 v63, s71, 23
; VI-NEXT:    v_writelane_b32 v63, s80, 24
; VI-NEXT:    v_writelane_b32 v63, s81, 25
; VI-NEXT:    v_writelane_b32 v63, s82, 26
; VI-NEXT:    v_writelane_b32 v63, s83, 27
; VI-NEXT:    v_writelane_b32 v63, s84, 28
; VI-NEXT:    v_writelane_b32 v63, s85, 29
; VI-NEXT:    v_writelane_b32 v63, s86, 30
; VI-NEXT:    v_readfirstlane_b32 s4, v19
; VI-NEXT:    v_writelane_b32 v63, s87, 31
; VI-NEXT:    v_readfirstlane_b32 s7, v18
; VI-NEXT:    v_readfirstlane_b32 s6, v17
; VI-NEXT:    v_readfirstlane_b32 s9, v16
; VI-NEXT:    v_readfirstlane_b32 s8, v15
; VI-NEXT:    v_readfirstlane_b32 s11, v14
; VI-NEXT:    v_readfirstlane_b32 s10, v13
; VI-NEXT:    v_readfirstlane_b32 s13, v12
; VI-NEXT:    v_readfirstlane_b32 s12, v11
; VI-NEXT:    v_readfirstlane_b32 s15, v10
; VI-NEXT:    v_readfirstlane_b32 s14, v9
; VI-NEXT:    v_readfirstlane_b32 s41, v8
; VI-NEXT:    v_readfirstlane_b32 s40, v7
; VI-NEXT:    v_readfirstlane_b32 s43, v6
; VI-NEXT:    v_readfirstlane_b32 s42, v5
; VI-NEXT:    v_readfirstlane_b32 s45, v4
; VI-NEXT:    v_readfirstlane_b32 s44, v3
; VI-NEXT:    v_readfirstlane_b32 s5, v2
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s4, v1
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; VI-NEXT:    s_cbranch_scc0 .LBB95_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s29, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s28, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s27, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s26, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s25, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s24, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s23, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s23, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s23, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s22, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s21, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s21, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s20, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s20, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s19, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s19, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s19, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s18, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s18, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s17, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s17, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 1
; VI-NEXT:    s_lshr_b32 s46, s17, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s16, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 0
; VI-NEXT:    s_lshr_b32 s46, s16, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v62, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s15, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s14, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s41, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s40, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s43, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s42, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v62, s46, 57
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    s_lshr_b32 s52, s7, 16
; VI-NEXT:    s_lshr_b32 s53, s6, 16
; VI-NEXT:    s_lshr_b32 s84, s9, 16
; VI-NEXT:    s_lshr_b32 s85, s8, 16
; VI-NEXT:    s_lshr_b32 s80, s11, 24
; VI-NEXT:    s_lshr_b32 s86, s11, 16
; VI-NEXT:    s_lshr_b32 s87, s10, 16
; VI-NEXT:    s_lshr_b32 s81, s13, 24
; VI-NEXT:    s_lshr_b32 s54, s13, 16
; VI-NEXT:    s_lshr_b32 s55, s12, 16
; VI-NEXT:    s_lshr_b32 s82, s15, 24
; VI-NEXT:    s_lshr_b32 s64, s15, 16
; VI-NEXT:    s_lshr_b32 s65, s14, 16
; VI-NEXT:    s_lshr_b32 s83, s41, 24
; VI-NEXT:    s_lshr_b32 s66, s41, 16
; VI-NEXT:    s_lshr_b32 s67, s40, 16
; VI-NEXT:    s_lshr_b32 s50, s43, 24
; VI-NEXT:    s_lshr_b32 s68, s43, 16
; VI-NEXT:    s_lshr_b32 s69, s42, 16
; VI-NEXT:    s_lshr_b32 s51, s45, 24
; VI-NEXT:    s_lshr_b32 s70, s45, 16
; VI-NEXT:    s_lshr_b32 s71, s44, 16
; VI-NEXT:    v_writelane_b32 v62, s46, 56
; VI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; VI-NEXT:    s_lshr_b64 s[56:57], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[16:17], 24
; VI-NEXT:    s_lshr_b64 s[78:79], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[42:43], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[44:45], 24
; VI-NEXT:    s_cbranch_execnz .LBB95_4
; VI-NEXT:  .LBB95_2: ; %cmp.true
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_mov_b32_e32 v7, 0x200
; VI-NEXT:    v_add_f16_e32 v58, s46, v7
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v58
; VI-NEXT:    v_add_f16_e32 v2, s45, v7
; VI-NEXT:    s_lshr_b32 s45, s44, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v23, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s45, v7
; VI-NEXT:    v_add_f16_e32 v2, s44, v7
; VI-NEXT:    s_lshr_b32 s44, s43, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v37, s44, v7
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v22, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v37
; VI-NEXT:    v_add_f16_e32 v2, s43, v7
; VI-NEXT:    s_lshr_b32 s43, s42, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v25, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s43, v7
; VI-NEXT:    v_add_f16_e32 v2, s42, v7
; VI-NEXT:    s_lshr_b32 s42, s41, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v50, s42, v7
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v24, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v50
; VI-NEXT:    v_add_f16_e32 v2, s41, v7
; VI-NEXT:    s_lshr_b32 s41, s40, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v27, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s41, v7
; VI-NEXT:    v_add_f16_e32 v2, s40, v7
; VI-NEXT:    s_lshr_b32 s40, s15, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v60, s40, v7
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v26, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v60
; VI-NEXT:    v_add_f16_e32 v2, s15, v7
; VI-NEXT:    s_lshr_b32 s15, s14, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v29, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s15, v7
; VI-NEXT:    v_add_f16_e32 v2, s14, v7
; VI-NEXT:    s_lshr_b32 s14, s13, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v61, s14, v7
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v28, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v61
; VI-NEXT:    v_add_f16_e32 v2, s13, v7
; VI-NEXT:    s_lshr_b32 s13, s12, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v6, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s13, v7
; VI-NEXT:    v_add_f16_e32 v2, s12, v7
; VI-NEXT:    s_lshr_b32 s12, s11, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v52, s12, v7
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v5, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v52
; VI-NEXT:    v_add_f16_e32 v2, s11, v7
; VI-NEXT:    s_lshr_b32 s11, s10, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v31, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s11, v7
; VI-NEXT:    v_add_f16_e32 v2, s10, v7
; VI-NEXT:    s_lshr_b32 s10, s9, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v53, s10, v7
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v30, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v53
; VI-NEXT:    v_add_f16_e32 v2, s9, v7
; VI-NEXT:    s_lshr_b32 s9, s8, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v4, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s9, v7
; VI-NEXT:    v_add_f16_e32 v2, s8, v7
; VI-NEXT:    s_lshr_b32 s8, s7, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v41, s8, v7
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v3, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v41
; VI-NEXT:    v_add_f16_e32 v2, s7, v7
; VI-NEXT:    s_lshr_b32 s7, s6, 16
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v2, v2, v1
; VI-NEXT:    v_add_f16_e32 v1, s7, v7
; VI-NEXT:    v_add_f16_e32 v8, s6, v7
; VI-NEXT:    s_lshr_b32 s6, s17, 16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_f16_e32 v43, s6, v7
; VI-NEXT:    s_lshr_b32 s6, s16, 16
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v1, v8, v1
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v43
; VI-NEXT:    v_add_f16_e32 v9, s17, v7
; VI-NEXT:    v_add_f16_e32 v11, s6, v7
; VI-NEXT:    s_lshr_b32 s6, s19, 16
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v33, v9, v8
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v11
; VI-NEXT:    v_add_f16_e32 v9, s16, v7
; VI-NEXT:    v_add_f16_e32 v36, s6, v7
; VI-NEXT:    s_lshr_b32 s6, s18, 16
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v32, v9, v8
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v36
; VI-NEXT:    v_add_f16_e32 v9, s19, v7
; VI-NEXT:    v_add_f16_e32 v38, s6, v7
; VI-NEXT:    s_lshr_b32 s6, s21, 16
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v21, v9, v8
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v38
; VI-NEXT:    v_add_f16_e32 v9, s18, v7
; VI-NEXT:    s_lshr_b32 s7, s20, 16
; VI-NEXT:    v_add_f16_e32 v49, s6, v7
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v20, v9, v8
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v49
; VI-NEXT:    v_add_f16_e32 v9, s21, v7
; VI-NEXT:    v_add_f16_e32 v39, s7, v7
; VI-NEXT:    s_lshr_b32 s6, s23, 16
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v35, v9, v8
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v39
; VI-NEXT:    v_add_f16_e32 v9, s20, v7
; VI-NEXT:    v_add_f16_e32 v45, s6, v7
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v34, v9, v8
; VI-NEXT:    s_lshr_b32 s7, s22, 16
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v45
; VI-NEXT:    v_add_f16_e32 v9, s23, v7
; VI-NEXT:    v_or_b32_e32 v19, v9, v8
; VI-NEXT:    v_add_f16_e32 v8, s7, v7
; VI-NEXT:    s_lshr_b32 s6, s25, 16
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_add_f16_e32 v9, s22, v7
; VI-NEXT:    v_add_f16_e32 v47, s6, v7
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v18, v9, v8
; VI-NEXT:    s_lshr_b32 s7, s24, 16
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v47
; VI-NEXT:    v_add_f16_e32 v9, s25, v7
; VI-NEXT:    v_or_b32_e32 v16, v9, v8
; VI-NEXT:    v_add_f16_e32 v8, s7, v7
; VI-NEXT:    s_lshr_b32 s6, s27, 16
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_add_f16_e32 v9, s24, v7
; VI-NEXT:    v_add_f16_e32 v57, s6, v7
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v15, v9, v8
; VI-NEXT:    s_lshr_b32 s7, s26, 16
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v57
; VI-NEXT:    v_add_f16_e32 v9, s27, v7
; VI-NEXT:    v_or_b32_e32 v13, v9, v8
; VI-NEXT:    v_add_f16_e32 v8, s7, v7
; VI-NEXT:    s_lshr_b32 s6, s29, 16
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_add_f16_e32 v54, s26, v7
; VI-NEXT:    v_add_f16_e32 v59, s6, v7
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v12, v54, v8
; VI-NEXT:    s_lshr_b32 s7, s28, 16
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v59
; VI-NEXT:    v_add_f16_e32 v9, s29, v7
; VI-NEXT:    v_or_b32_e32 v10, v9, v8
; VI-NEXT:    v_add_f16_e32 v8, s7, v7
; VI-NEXT:    s_lshr_b32 s6, s5, 16
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_add_f16_e32 v55, s28, v7
; VI-NEXT:    s_lshr_b32 s7, s4, 16
; VI-NEXT:    v_add_f16_e32 v51, s6, v7
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v9, v55, v8
; VI-NEXT:    v_add_f16_e32 v8, s5, v7
; VI-NEXT:    v_add_f16_e32 v14, s7, v7
; VI-NEXT:    v_add_f16_e32 v17, s4, v7
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v51
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v8, v8, v7
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v14
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v7, v17, v7
; VI-NEXT:    v_lshrrev_b32_e32 v14, 8, v8
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v14, 8, v7
; VI-NEXT:    v_lshrrev_b64 v[7:8], 24, v[7:8]
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v10
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v9
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[7:8], 24, v[9:10]
; VI-NEXT:    v_lshrrev_b32_e32 v8, 8, v13
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v8, 8, v12
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v8, 8, v16
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v8, 8, v15
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[16:17], 24, v[15:16]
; VI-NEXT:    v_lshrrev_b32_e32 v8, 8, v19
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v10, 8, v18
; VI-NEXT:    v_lshrrev_b64 v[17:18], 24, v[18:19]
; VI-NEXT:    v_lshrrev_b32_e32 v8, 8, v2
; VI-NEXT:    v_lshrrev_b32_e32 v48, 8, v35
; VI-NEXT:    v_lshrrev_b32_e32 v44, 8, v34
; VI-NEXT:    v_lshrrev_b64 v[18:19], 24, v[34:35]
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v8, 8, v1
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[1:2]
; VI-NEXT:    v_lshrrev_b32_e32 v34, 8, v25
; VI-NEXT:    v_lshrrev_b32_e32 v35, 8, v24
; VI-NEXT:    v_lshrrev_b64 v[24:25], 24, v[24:25]
; VI-NEXT:    v_bfe_u32 v25, v41, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v25, v53, 8, 8
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v4
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v25, v52, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v3
; VI-NEXT:    v_lshrrev_b64 v[2:3], 24, v[3:4]
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v25, v61, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v31
; VI-NEXT:    v_lshrrev_b64 v[3:4], 24, v[30:31]
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v25, v60, 8, 8
; VI-NEXT:    v_lshrrev_b32_e32 v46, 8, v20
; VI-NEXT:    v_lshrrev_b64 v[19:20], 24, v[20:21]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v30
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v5
; VI-NEXT:    v_lshrrev_b64 v[4:5], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v25, v50, 8, 8
; VI-NEXT:    v_lshrrev_b32_e32 v42, 8, v21
; VI-NEXT:    v_lshrrev_b32_e32 v56, 8, v32
; VI-NEXT:    v_lshrrev_b64 v[20:21], 24, v[32:33]
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v6
; VI-NEXT:    v_lshrrev_b32_e32 v32, 8, v29
; VI-NEXT:    v_lshrrev_b64 v[5:6], 24, v[28:29]
; VI-NEXT:    v_lshrrev_b32_e32 v29, 8, v27
; VI-NEXT:    v_lshrrev_b64 v[8:9], 24, v[26:27]
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v25, v37, 8, 8
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v27, v58, 8, 8
; VI-NEXT:    v_mov_b32_e32 v58, v37
; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[13:14], 24, v[12:13]
; VI-NEXT:    v_lshrrev_b64 v[14:15], 24, v[22:23]
; VI-NEXT:    v_lshrrev_b32_e32 v40, 8, v33
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v33, 8, v28
; VI-NEXT:    v_lshrrev_b32_e32 v6, 8, v26
; VI-NEXT:    v_lshrrev_b32_e32 v21, 8, v23
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v22
; VI-NEXT:    v_mov_b32_e32 v23, v11
; VI-NEXT:    v_bfe_u32 v11, v51, 8, 8
; VI-NEXT:    v_bfe_u32 v26, v59, 8, 8
; VI-NEXT:    v_bfe_u32 v28, v57, 8, 8
; VI-NEXT:    v_bfe_u32 v9, v47, 8, 8
; VI-NEXT:    v_bfe_u32 v12, v45, 8, 8
; VI-NEXT:    v_bfe_u32 v15, v49, 8, 8
; VI-NEXT:    v_bfe_u32 v22, v36, 8, 8
; VI-NEXT:    v_bfe_u32 v30, v43, 8, 8
; VI-NEXT:    s_branch .LBB95_5
; VI-NEXT:  .LBB95_3:
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr71
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr51
; VI-NEXT:    ; implicit-def: $sgpr69
; VI-NEXT:    ; implicit-def: $sgpr68
; VI-NEXT:    ; implicit-def: $sgpr50
; VI-NEXT:    ; implicit-def: $sgpr67
; VI-NEXT:    ; implicit-def: $sgpr66
; VI-NEXT:    ; implicit-def: $sgpr83
; VI-NEXT:    ; implicit-def: $sgpr65
; VI-NEXT:    ; implicit-def: $sgpr64
; VI-NEXT:    ; implicit-def: $sgpr82
; VI-NEXT:    ; implicit-def: $sgpr55
; VI-NEXT:    ; implicit-def: $sgpr54
; VI-NEXT:    ; implicit-def: $sgpr81
; VI-NEXT:    ; implicit-def: $sgpr87
; VI-NEXT:    ; implicit-def: $sgpr86
; VI-NEXT:    ; implicit-def: $sgpr80
; VI-NEXT:    ; implicit-def: $sgpr85
; VI-NEXT:    ; implicit-def: $sgpr84
; VI-NEXT:    ; implicit-def: $sgpr53
; VI-NEXT:    ; implicit-def: $sgpr52
; VI-NEXT:    ; implicit-def: $sgpr76
; VI-NEXT:    ; implicit-def: $sgpr74
; VI-NEXT:    ; implicit-def: $sgpr72
; VI-NEXT:    ; implicit-def: $sgpr62
; VI-NEXT:    ; implicit-def: $sgpr60
; VI-NEXT:    ; implicit-def: $sgpr58
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr48
; VI-NEXT:    ; implicit-def: $sgpr38
; VI-NEXT:    ; implicit-def: $sgpr36
; VI-NEXT:    ; implicit-def: $sgpr34
; VI-NEXT:    ; implicit-def: $sgpr30
; VI-NEXT:    ; implicit-def: $sgpr90
; VI-NEXT:    ; implicit-def: $sgpr88
; VI-NEXT:    ; implicit-def: $sgpr78
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    s_branch .LBB95_2
; VI-NEXT:  .LBB95_4:
; VI-NEXT:    v_mov_b32_e32 v1, s44
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s45
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s42
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s43
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s40
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s41
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s14
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s15
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s12
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s13
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s10
; VI-NEXT:    v_mov_b32_e32 v2, s46
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s11
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s9
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s7
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s16
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s18
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s20
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s21
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s22
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s23
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s24
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s25
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s27
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s29
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s70
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s69
; VI-NEXT:    v_readlane_b32 s6, v62, 0
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s67
; VI-NEXT:    v_mov_b32_e32 v23, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 1
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s65
; VI-NEXT:    v_mov_b32_e32 v43, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 2
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s55
; VI-NEXT:    v_mov_b32_e32 v38, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 3
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s87
; VI-NEXT:    v_mov_b32_e32 v36, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 4
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s85
; VI-NEXT:    v_mov_b32_e32 v39, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 5
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s53
; VI-NEXT:    v_mov_b32_e32 v49, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 6
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 7
; VI-NEXT:    v_mov_b32_e32 v45, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 9
; VI-NEXT:    v_mov_b32_e32 v47, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 10
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 11
; VI-NEXT:    v_mov_b32_e32 v57, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 12
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 13
; VI-NEXT:    v_mov_b32_e32 v59, s6
; VI-NEXT:    v_readlane_b32 s6, v62, 14
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 16
; VI-NEXT:    v_mov_b32_e32 v11, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 17
; VI-NEXT:    v_mov_b32_e32 v26, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 18
; VI-NEXT:    v_mov_b32_e32 v28, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 19
; VI-NEXT:    v_mov_b32_e32 v9, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 20
; VI-NEXT:    v_mov_b32_e32 v12, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 21
; VI-NEXT:    v_mov_b32_e32 v15, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 22
; VI-NEXT:    v_mov_b32_e32 v22, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 23
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s5
; VI-NEXT:    v_mov_b32_e32 v30, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 24
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 25
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s80
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s81
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s82
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s83
; VI-NEXT:    v_readlane_b32 s4, v62, 26
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 27
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 28
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 29
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 30
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 31
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 32
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 33
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 34
; VI-NEXT:    v_mov_b32_e32 v10, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 35
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 36
; VI-NEXT:    v_mov_b32_e32 v44, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 37
; VI-NEXT:    v_mov_b32_e32 v48, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 38
; VI-NEXT:    v_mov_b32_e32 v46, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 39
; VI-NEXT:    v_mov_b32_e32 v42, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 40
; VI-NEXT:    v_mov_b32_e32 v56, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 41
; VI-NEXT:    v_mov_b32_e32 v40, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 42
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 43
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 44
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 45
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 46
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 47
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 48
; VI-NEXT:    v_mov_b32_e32 v31, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 49
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 50
; VI-NEXT:    v_mov_b32_e32 v33, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 51
; VI-NEXT:    v_mov_b32_e32 v32, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 52
; VI-NEXT:    v_mov_b32_e32 v6, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 53
; VI-NEXT:    v_mov_b32_e32 v29, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 54
; VI-NEXT:    v_mov_b32_e32 v35, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 55
; VI-NEXT:    v_mov_b32_e32 v34, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 56
; VI-NEXT:    v_mov_b32_e32 v2, s78
; VI-NEXT:    v_readlane_b32 s6, v62, 15
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_readlane_b32 s4, v62, 57
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v54, s26
; VI-NEXT:    v_mov_b32_e32 v55, s28
; VI-NEXT:    v_mov_b32_e32 v37, s71
; VI-NEXT:    v_mov_b32_e32 v58, s68
; VI-NEXT:    v_mov_b32_e32 v50, s66
; VI-NEXT:    v_mov_b32_e32 v60, s64
; VI-NEXT:    v_mov_b32_e32 v61, s54
; VI-NEXT:    v_mov_b32_e32 v52, s86
; VI-NEXT:    v_mov_b32_e32 v53, s84
; VI-NEXT:    v_mov_b32_e32 v41, s52
; VI-NEXT:    v_mov_b32_e32 v51, s6
; VI-NEXT:    v_mov_b32_e32 v25, s50
; VI-NEXT:    v_mov_b32_e32 v27, s51
; VI-NEXT:    v_mov_b32_e32 v21, s4
; VI-NEXT:    v_mov_b32_e32 v7, s56
; VI-NEXT:    v_mov_b32_e32 v13, s58
; VI-NEXT:    v_mov_b32_e32 v16, s60
; VI-NEXT:    v_mov_b32_e32 v17, s62
; VI-NEXT:    v_mov_b32_e32 v18, s72
; VI-NEXT:    v_mov_b32_e32 v19, s74
; VI-NEXT:    v_mov_b32_e32 v20, s76
; VI-NEXT:    v_mov_b32_e32 v2, s88
; VI-NEXT:    v_mov_b32_e32 v3, s90
; VI-NEXT:    v_mov_b32_e32 v4, s30
; VI-NEXT:    v_mov_b32_e32 v5, s34
; VI-NEXT:    v_mov_b32_e32 v8, s36
; VI-NEXT:    v_mov_b32_e32 v24, s38
; VI-NEXT:    v_mov_b32_e32 v14, s48
; VI-NEXT:  .LBB95_5: ; %end
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_perm_b32 v20, v23, v20, s4
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_perm_b32 v30, v43, v30, s4
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_perm_b32 v19, v38, v19, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_perm_b32 v18, v39, v18, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v15, v49, v15, s4
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v12, v45, v12, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v9, v47, v9, s4
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_readlane_b32 s87, v63, 31
; VI-NEXT:    v_readlane_b32 s86, v63, 30
; VI-NEXT:    v_readlane_b32 s85, v63, 29
; VI-NEXT:    v_readlane_b32 s84, v63, 28
; VI-NEXT:    v_readlane_b32 s83, v63, 27
; VI-NEXT:    v_readlane_b32 s82, v63, 26
; VI-NEXT:    v_readlane_b32 s81, v63, 25
; VI-NEXT:    v_readlane_b32 s80, v63, 24
; VI-NEXT:    v_readlane_b32 s71, v63, 23
; VI-NEXT:    v_readlane_b32 s70, v63, 22
; VI-NEXT:    v_readlane_b32 s69, v63, 21
; VI-NEXT:    v_readlane_b32 s68, v63, 20
; VI-NEXT:    v_readlane_b32 s67, v63, 19
; VI-NEXT:    v_readlane_b32 s66, v63, 18
; VI-NEXT:    v_readlane_b32 s65, v63, 17
; VI-NEXT:    v_readlane_b32 s64, v63, 16
; VI-NEXT:    v_readlane_b32 s55, v63, 15
; VI-NEXT:    v_readlane_b32 s54, v63, 14
; VI-NEXT:    v_readlane_b32 s53, v63, 13
; VI-NEXT:    v_readlane_b32 s52, v63, 12
; VI-NEXT:    v_readlane_b32 s51, v63, 11
; VI-NEXT:    v_readlane_b32 s50, v63, 10
; VI-NEXT:    v_readlane_b32 s49, v63, 9
; VI-NEXT:    v_readlane_b32 s48, v63, 8
; VI-NEXT:    v_readlane_b32 s39, v63, 7
; VI-NEXT:    v_readlane_b32 s38, v63, 6
; VI-NEXT:    v_readlane_b32 s37, v63, 5
; VI-NEXT:    v_readlane_b32 s36, v63, 4
; VI-NEXT:    v_readlane_b32 s35, v63, 3
; VI-NEXT:    v_readlane_b32 s34, v63, 2
; VI-NEXT:    v_readlane_b32 s31, v63, 1
; VI-NEXT:    v_readlane_b32 s30, v63, 0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v56, v23, v56, s4
; VI-NEXT:    v_or_b32_e32 v20, v56, v20
; VI-NEXT:    buffer_store_dword v20, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v20, v40, s4
; VI-NEXT:    v_or_b32_e32 v20, v20, v30
; VI-NEXT:    v_add_u32_e32 v30, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v20, v30, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v20, v46, s4
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_add_u32_e32 v20, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v19, v20, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v20, v36, v22, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v19, v42, s4
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    v_add_u32_e32 v20, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v19, v20, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v19, v44, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_add_u32_e32 v19, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v18, v19, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v18, v48, s4
; VI-NEXT:    v_or_b32_e32 v15, v18, v15
; VI-NEXT:    v_add_u32_e32 v18, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v15, v18, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v15, v17, s4
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v17, v10, s4
; VI-NEXT:    v_or_b32_e32 v15, v17, v15
; VI-NEXT:    v_add_u32_e32 v17, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v15, v17, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v15, v10, s4
; VI-NEXT:    v_or_b32_e32 v12, v15, v12
; VI-NEXT:    v_add_u32_e32 v15, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v12, v15, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v10, v16, s4
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v15, v10, s4
; VI-NEXT:    v_or_b32_e32 v12, v15, v12
; VI-NEXT:    v_add_u32_e32 v15, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v12, v15, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v12, v10, s4
; VI-NEXT:    v_or_b32_e32 v9, v12, v9
; VI-NEXT:    v_add_u32_e32 v12, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v9, v12, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v9, v9, v13, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v54, v10, s4
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_e32 v9, v12, v9
; VI-NEXT:    v_add_u32_e32 v12, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v9, v12, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v12, v57, v28, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v9, v10, s4
; VI-NEXT:    v_or_b32_e32 v9, v9, v12
; VI-NEXT:    v_add_u32_e32 v12, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v9, v12, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v9, v7, s4
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v10, v55, v7, s4
; VI-NEXT:    v_or_b32_e32 v9, v10, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v9, v10, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v10, v59, v26, s4
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v9, v7, s4
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_add_u32_e32 v10, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v9, v10, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v7, v7, v9, s4
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v9, v9, v10, s4
; VI-NEXT:    v_or_b32_e32 v7, v9, v7
; VI-NEXT:    v_add_u32_e32 v9, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v7, v9, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v7, v9, s4
; VI-NEXT:    v_perm_b32 v9, v51, v11, s4
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_e32 v7, v7, v9
; VI-NEXT:    v_add_u32_e32 v9, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v7, v9, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v7, v37, v14, s4
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v9, v1, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v7
; VI-NEXT:    v_add_u32_e32 v7, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v7, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v21, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v7, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v1, v1, v7
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v7, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v24, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v7, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v7, v1
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v7, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v7, v58, v25, s4
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v34, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v7
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v7, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v6, s4
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v6, v8, s4
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v1, v1, v6
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v6, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v29, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v50, v6, s4
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v1, v1, v6
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v6, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v33, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v1, v1, v5
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v5, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v32, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v60, v5, s4
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v1, v1, v5
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v5, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v31, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s4
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v1, v1, v4
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v4, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v4, s4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v61, v4, s4
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v1, v1, v4
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v4, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v4, s4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v1, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v3, s4
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v52, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v1, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v3, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v3, s4
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v53, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v41, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v128i8_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v63, s30, 0
; GFX9-NEXT:    v_writelane_b32 v63, s31, 1
; GFX9-NEXT:    v_writelane_b32 v63, s34, 2
; GFX9-NEXT:    v_writelane_b32 v63, s35, 3
; GFX9-NEXT:    v_writelane_b32 v63, s36, 4
; GFX9-NEXT:    v_writelane_b32 v63, s37, 5
; GFX9-NEXT:    v_writelane_b32 v63, s38, 6
; GFX9-NEXT:    v_writelane_b32 v63, s39, 7
; GFX9-NEXT:    v_writelane_b32 v63, s48, 8
; GFX9-NEXT:    v_writelane_b32 v63, s49, 9
; GFX9-NEXT:    v_writelane_b32 v63, s50, 10
; GFX9-NEXT:    v_writelane_b32 v63, s51, 11
; GFX9-NEXT:    v_writelane_b32 v63, s52, 12
; GFX9-NEXT:    v_writelane_b32 v63, s53, 13
; GFX9-NEXT:    v_writelane_b32 v63, s54, 14
; GFX9-NEXT:    v_writelane_b32 v63, s55, 15
; GFX9-NEXT:    v_writelane_b32 v63, s64, 16
; GFX9-NEXT:    v_writelane_b32 v63, s65, 17
; GFX9-NEXT:    v_writelane_b32 v63, s66, 18
; GFX9-NEXT:    v_writelane_b32 v63, s67, 19
; GFX9-NEXT:    v_writelane_b32 v63, s68, 20
; GFX9-NEXT:    v_writelane_b32 v63, s69, 21
; GFX9-NEXT:    v_writelane_b32 v63, s70, 22
; GFX9-NEXT:    v_writelane_b32 v63, s71, 23
; GFX9-NEXT:    v_writelane_b32 v63, s80, 24
; GFX9-NEXT:    v_writelane_b32 v63, s81, 25
; GFX9-NEXT:    v_writelane_b32 v63, s82, 26
; GFX9-NEXT:    v_writelane_b32 v63, s83, 27
; GFX9-NEXT:    v_writelane_b32 v63, s84, 28
; GFX9-NEXT:    v_writelane_b32 v63, s85, 29
; GFX9-NEXT:    v_writelane_b32 v63, s86, 30
; GFX9-NEXT:    v_writelane_b32 v63, s87, 31
; GFX9-NEXT:    v_writelane_b32 v63, s96, 32
; GFX9-NEXT:    v_writelane_b32 v63, s97, 33
; GFX9-NEXT:    v_writelane_b32 v63, s98, 34
; GFX9-NEXT:    v_readfirstlane_b32 s4, v19
; GFX9-NEXT:    v_writelane_b32 v63, s99, 35
; GFX9-NEXT:    v_readfirstlane_b32 s7, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s9, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s11, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s13, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s15, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s41, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s43, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s45, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s5, v2
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s4, v1
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; GFX9-NEXT:    s_cbranch_scc0 .LBB95_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 49
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s29, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s28, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s27, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s27, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s26, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s26, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s25, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s25, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s25, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s24, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s24, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s23, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s23, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s23, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s22, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s22, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s21, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s21, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s21, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s20, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s20, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s19, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s19, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s19, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s18, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s18, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s17, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s17, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s17, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s16, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s16, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s6, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s6, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s8, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s8, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 0
; GFX9-NEXT:    s_lshr_b32 s82, s11, 24
; GFX9-NEXT:    s_lshr_b32 s83, s11, 16
; GFX9-NEXT:    s_lshr_b32 s85, s11, 8
; GFX9-NEXT:    s_lshr_b32 s84, s10, 16
; GFX9-NEXT:    s_lshr_b32 s86, s10, 8
; GFX9-NEXT:    s_lshr_b32 s87, s13, 24
; GFX9-NEXT:    s_lshr_b32 s96, s13, 16
; GFX9-NEXT:    s_lshr_b32 s98, s13, 8
; GFX9-NEXT:    s_lshr_b32 s97, s12, 16
; GFX9-NEXT:    s_lshr_b32 s99, s12, 8
; GFX9-NEXT:    s_lshr_b32 s38, s15, 24
; GFX9-NEXT:    s_lshr_b32 s39, s15, 16
; GFX9-NEXT:    s_lshr_b32 s49, s15, 8
; GFX9-NEXT:    s_lshr_b32 s48, s14, 16
; GFX9-NEXT:    s_lshr_b32 s50, s14, 8
; GFX9-NEXT:    s_lshr_b32 s51, s41, 24
; GFX9-NEXT:    s_lshr_b32 s52, s41, 16
; GFX9-NEXT:    s_lshr_b32 s54, s41, 8
; GFX9-NEXT:    s_lshr_b32 s53, s40, 16
; GFX9-NEXT:    s_lshr_b32 s55, s40, 8
; GFX9-NEXT:    s_lshr_b32 s64, s43, 24
; GFX9-NEXT:    s_lshr_b32 s65, s43, 16
; GFX9-NEXT:    s_lshr_b32 s67, s43, 8
; GFX9-NEXT:    s_lshr_b32 s66, s42, 16
; GFX9-NEXT:    s_lshr_b32 s68, s42, 8
; GFX9-NEXT:    s_lshr_b32 s69, s45, 24
; GFX9-NEXT:    s_lshr_b32 s70, s45, 16
; GFX9-NEXT:    s_lshr_b32 s80, s45, 8
; GFX9-NEXT:    s_lshr_b32 s71, s44, 16
; GFX9-NEXT:    s_lshr_b32 s81, s44, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[16:17], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[44:45], 24
; GFX9-NEXT:    s_cbranch_execnz .LBB95_4
; GFX9-NEXT:  .LBB95_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v15, 0x200
; GFX9-NEXT:    v_pk_add_f16 v26, s5, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, s4, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, s45, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, s44, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, s43, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, s42, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, s41, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, s40, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, s15, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, s14, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, s13, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, s12, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, s11, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, s10, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, s9, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, s8, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, s7, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, s6, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v49, s17, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v48, s16, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v38, s19, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v37, s18, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v36, s21, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v35, s20, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v34, s23, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v33, s22, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v32, s25, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v31, s24, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, s27, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, s26, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, s29, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, s28, v15 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[33:34]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[35:36]
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v38
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v38
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v49
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v49
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v4
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v3
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v3
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v6
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v6
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v5
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v5
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[9:10]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v8
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v8
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v8
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v7
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v7
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v10
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v10
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v10
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v9
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v9
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v26
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v26
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v26
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v25
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v28
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v28
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v30
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v30
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v32
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v20
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v32
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v20
; GFX9-NEXT:    v_lshrrev_b64 v[43:44], 24, v[37:38]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 24, v36
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v20
; GFX9-NEXT:    v_lshrrev_b64 v[44:45], 24, v[48:49]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v34
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 16, v36
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 16, v25
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 8, v28
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 16, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v30
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v29
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v29
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 16, v31
; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 8, v31
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v23, 8, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v33
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v33
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 8, v36
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v35
; GFX9-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
; GFX9-NEXT:    v_lshrrev_b32_e32 v24, 8, v38
; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 16, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 16, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 8, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v40, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v41, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v1
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v19
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    s_branch .LBB95_5
; GFX9-NEXT:  .LBB95_3:
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr81
; GFX9-NEXT:    ; implicit-def: $sgpr71
; GFX9-NEXT:    ; implicit-def: $sgpr80
; GFX9-NEXT:    ; implicit-def: $sgpr70
; GFX9-NEXT:    ; implicit-def: $sgpr69
; GFX9-NEXT:    ; implicit-def: $sgpr68
; GFX9-NEXT:    ; implicit-def: $sgpr66
; GFX9-NEXT:    ; implicit-def: $sgpr67
; GFX9-NEXT:    ; implicit-def: $sgpr65
; GFX9-NEXT:    ; implicit-def: $sgpr64
; GFX9-NEXT:    ; implicit-def: $sgpr55
; GFX9-NEXT:    ; implicit-def: $sgpr53
; GFX9-NEXT:    ; implicit-def: $sgpr54
; GFX9-NEXT:    ; implicit-def: $sgpr52
; GFX9-NEXT:    ; implicit-def: $sgpr51
; GFX9-NEXT:    ; implicit-def: $sgpr50
; GFX9-NEXT:    ; implicit-def: $sgpr48
; GFX9-NEXT:    ; implicit-def: $sgpr49
; GFX9-NEXT:    ; implicit-def: $sgpr39
; GFX9-NEXT:    ; implicit-def: $sgpr38
; GFX9-NEXT:    ; implicit-def: $sgpr99
; GFX9-NEXT:    ; implicit-def: $sgpr97
; GFX9-NEXT:    ; implicit-def: $sgpr98
; GFX9-NEXT:    ; implicit-def: $sgpr96
; GFX9-NEXT:    ; implicit-def: $sgpr87
; GFX9-NEXT:    ; implicit-def: $sgpr86
; GFX9-NEXT:    ; implicit-def: $sgpr84
; GFX9-NEXT:    ; implicit-def: $sgpr85
; GFX9-NEXT:    ; implicit-def: $sgpr83
; GFX9-NEXT:    ; implicit-def: $sgpr82
; GFX9-NEXT:    ; implicit-def: $sgpr76
; GFX9-NEXT:    ; implicit-def: $sgpr74
; GFX9-NEXT:    ; implicit-def: $sgpr72
; GFX9-NEXT:    ; implicit-def: $sgpr62
; GFX9-NEXT:    ; implicit-def: $sgpr60
; GFX9-NEXT:    ; implicit-def: $sgpr58
; GFX9-NEXT:    ; implicit-def: $sgpr56
; GFX9-NEXT:    ; implicit-def: $sgpr36
; GFX9-NEXT:    ; implicit-def: $sgpr34
; GFX9-NEXT:    ; implicit-def: $sgpr30
; GFX9-NEXT:    ; implicit-def: $sgpr94
; GFX9-NEXT:    ; implicit-def: $sgpr92
; GFX9-NEXT:    ; implicit-def: $sgpr90
; GFX9-NEXT:    ; implicit-def: $sgpr88
; GFX9-NEXT:    ; implicit-def: $sgpr78
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    s_branch .LBB95_2
; GFX9-NEXT:  .LBB95_4:
; GFX9-NEXT:    v_mov_b32_e32 v15, s81
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s71
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s80
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s70
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s69
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s68
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s66
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s67
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s65
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s64
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s55
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s53
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s54
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s52
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s50
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s48
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s49
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s39
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s38
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s99
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s97
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s98
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s96
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s87
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s86
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s84
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s85
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s83
; GFX9-NEXT:    v_mov_b32_e32 v25, s4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s82
; GFX9-NEXT:    v_readlane_b32 s4, v62, 0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 3
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 5
; GFX9-NEXT:    v_mov_b32_e32 v22, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 7
; GFX9-NEXT:    v_mov_b32_e32 v45, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 8
; GFX9-NEXT:    v_mov_b32_e32 v41, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 9
; GFX9-NEXT:    v_mov_b32_e32 v42, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 10
; GFX9-NEXT:    v_mov_b32_e32 v55, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 11
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 12
; GFX9-NEXT:    v_mov_b32_e32 v50, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 15
; GFX9-NEXT:    v_mov_b32_e32 v52, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 16
; GFX9-NEXT:    v_mov_b32_e32 v54, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 17
; GFX9-NEXT:    v_mov_b32_e32 v24, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 18
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 19
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 20
; GFX9-NEXT:    v_mov_b32_e32 v21, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 21
; GFX9-NEXT:    v_mov_b32_e32 v18, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 22
; GFX9-NEXT:    v_mov_b32_e32 v16, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 23
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 25
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 26
; GFX9-NEXT:    v_mov_b32_e32 v61, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 27
; GFX9-NEXT:    v_mov_b32_e32 v23, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 28
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 29
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 30
; GFX9-NEXT:    v_mov_b32_e32 v53, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 31
; GFX9-NEXT:    v_mov_b32_e32 v60, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 32
; GFX9-NEXT:    v_mov_b32_e32 v59, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 33
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 34
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 35
; GFX9-NEXT:    v_mov_b32_e32 v57, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 36
; GFX9-NEXT:    v_mov_b32_e32 v58, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 37
; GFX9-NEXT:    v_mov_b32_e32 v56, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 38
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 39
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 40
; GFX9-NEXT:    v_mov_b32_e32 v46, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 41
; GFX9-NEXT:    v_mov_b32_e32 v47, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 42
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 43
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 44
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 45
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 46
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 47
; GFX9-NEXT:    v_mov_b32_e32 v40, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 48
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 49
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s4
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s46
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s56
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s58
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s60
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s62
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s72
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s78
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s88
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v19, s44
; GFX9-NEXT:    v_mov_b32_e32 v20, s45
; GFX9-NEXT:    v_mov_b32_e32 v13, s42
; GFX9-NEXT:    v_mov_b32_e32 v14, s43
; GFX9-NEXT:    v_mov_b32_e32 v11, s40
; GFX9-NEXT:    v_mov_b32_e32 v12, s41
; GFX9-NEXT:    v_mov_b32_e32 v9, s14
; GFX9-NEXT:    v_mov_b32_e32 v10, s15
; GFX9-NEXT:    v_mov_b32_e32 v7, s12
; GFX9-NEXT:    v_mov_b32_e32 v8, s13
; GFX9-NEXT:    v_mov_b32_e32 v5, s10
; GFX9-NEXT:    v_mov_b32_e32 v6, s11
; GFX9-NEXT:    v_mov_b32_e32 v3, s8
; GFX9-NEXT:    v_mov_b32_e32 v4, s9
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    v_mov_b32_e32 v2, s7
; GFX9-NEXT:    v_mov_b32_e32 v48, s16
; GFX9-NEXT:    v_mov_b32_e32 v49, s17
; GFX9-NEXT:    v_mov_b32_e32 v37, s18
; GFX9-NEXT:    v_mov_b32_e32 v38, s19
; GFX9-NEXT:    v_mov_b32_e32 v35, s20
; GFX9-NEXT:    v_mov_b32_e32 v36, s21
; GFX9-NEXT:    v_mov_b32_e32 v33, s22
; GFX9-NEXT:    v_mov_b32_e32 v34, s23
; GFX9-NEXT:    v_mov_b32_e32 v31, s24
; GFX9-NEXT:    v_mov_b32_e32 v32, s25
; GFX9-NEXT:    v_mov_b32_e32 v29, s26
; GFX9-NEXT:    v_mov_b32_e32 v30, s27
; GFX9-NEXT:    v_mov_b32_e32 v27, s28
; GFX9-NEXT:    v_mov_b32_e32 v28, s29
; GFX9-NEXT:    v_mov_b32_e32 v26, s5
; GFX9-NEXT:    v_mov_b32_e32 v43, s74
; GFX9-NEXT:    v_mov_b32_e32 v44, s76
; GFX9-NEXT:    v_mov_b32_e32 v40, v41
; GFX9-NEXT:    v_mov_b32_e32 v41, v45
; GFX9-NEXT:    v_mov_b32_e32 v45, v22
; GFX9-NEXT:    v_mov_b32_e32 v22, s90
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s92
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s94
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s30
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s34
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s36
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:  .LBB95_5: ; %end
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v37, v37, v52, s4
; GFX9-NEXT:    v_perm_b32 v31, v31, v53, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v33, v33, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v21, v35, v21, s4
; GFX9-NEXT:    v_perm_b32 v16, v36, v16, s4
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v24, v38, v24, s4
; GFX9-NEXT:    v_perm_b32 v17, v28, v17, s4
; GFX9-NEXT:    v_perm_b32 v48, v48, v55, s4
; GFX9-NEXT:    v_perm_b32 v49, v49, v50, s4
; GFX9-NEXT:    v_perm_b32 v50, v54, v43, s4
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v39, v39, v44, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX9-NEXT:    v_or_b32_e32 v39, v48, v39
; GFX9-NEXT:    v_perm_b32 v23, v34, v23, s4
; GFX9-NEXT:    v_perm_b32 v32, v32, v59, s4
; GFX9-NEXT:    v_perm_b32 v29, v29, v57, s4
; GFX9-NEXT:    v_perm_b32 v30, v30, v56, s4
; GFX9-NEXT:    v_perm_b32 v27, v27, v46, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v45, s4
; GFX9-NEXT:    v_readlane_b32 s99, v63, 35
; GFX9-NEXT:    v_readlane_b32 s98, v63, 34
; GFX9-NEXT:    v_readlane_b32 s97, v63, 33
; GFX9-NEXT:    v_readlane_b32 s96, v63, 32
; GFX9-NEXT:    v_readlane_b32 s87, v63, 31
; GFX9-NEXT:    v_readlane_b32 s86, v63, 30
; GFX9-NEXT:    v_readlane_b32 s85, v63, 29
; GFX9-NEXT:    v_readlane_b32 s84, v63, 28
; GFX9-NEXT:    v_readlane_b32 s83, v63, 27
; GFX9-NEXT:    v_readlane_b32 s82, v63, 26
; GFX9-NEXT:    v_readlane_b32 s81, v63, 25
; GFX9-NEXT:    v_readlane_b32 s80, v63, 24
; GFX9-NEXT:    v_readlane_b32 s71, v63, 23
; GFX9-NEXT:    v_readlane_b32 s70, v63, 22
; GFX9-NEXT:    v_readlane_b32 s69, v63, 21
; GFX9-NEXT:    v_readlane_b32 s68, v63, 20
; GFX9-NEXT:    v_readlane_b32 s67, v63, 19
; GFX9-NEXT:    v_readlane_b32 s66, v63, 18
; GFX9-NEXT:    v_readlane_b32 s65, v63, 17
; GFX9-NEXT:    v_readlane_b32 s64, v63, 16
; GFX9-NEXT:    v_readlane_b32 s55, v63, 15
; GFX9-NEXT:    v_readlane_b32 s54, v63, 14
; GFX9-NEXT:    v_readlane_b32 s53, v63, 13
; GFX9-NEXT:    v_readlane_b32 s52, v63, 12
; GFX9-NEXT:    v_readlane_b32 s51, v63, 11
; GFX9-NEXT:    v_readlane_b32 s50, v63, 10
; GFX9-NEXT:    v_readlane_b32 s49, v63, 9
; GFX9-NEXT:    v_readlane_b32 s48, v63, 8
; GFX9-NEXT:    v_readlane_b32 s39, v63, 7
; GFX9-NEXT:    v_readlane_b32 s38, v63, 6
; GFX9-NEXT:    v_readlane_b32 s37, v63, 5
; GFX9-NEXT:    v_readlane_b32 s36, v63, 4
; GFX9-NEXT:    v_readlane_b32 s35, v63, 3
; GFX9-NEXT:    v_readlane_b32 s34, v63, 2
; GFX9-NEXT:    v_readlane_b32 s31, v63, 1
; GFX9-NEXT:    v_readlane_b32 s30, v63, 0
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v36, v58, v52, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v22, v25, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v35, v61, v35, s4
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v18, v18, v54, s4
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_or_b32_e32 v18, v21, v18
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v38, v47, v52, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v25, v26, v15, s4
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v34, v60, v54, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v28, v51, v52, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v15, v51, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v19, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v20, v15, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v15, v15, v51, s4
; GFX9-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v39, v39, v48, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX9-NEXT:    v_or_b32_e32 v39, v49, v39
; GFX9-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v50
; GFX9-NEXT:    v_or_b32_e32 v37, v37, v39
; GFX9-NEXT:    buffer_store_dword v37, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v37, v37, v39, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX9-NEXT:    v_or_b32_e32 v24, v24, v37
; GFX9-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v21, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v18
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v35
; GFX9-NEXT:    v_or_b32_e32 v16, v33, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v23, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v34
; GFX9-NEXT:    v_or_b32_e32 v16, v31, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v32, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v36
; GFX9-NEXT:    v_or_b32_e32 v16, v29, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v30, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v38
; GFX9-NEXT:    v_or_b32_e32 v16, v27, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v28
; GFX9-NEXT:    v_or_b32_e32 v16, v22, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v25, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v26
; GFX9-NEXT:    v_or_b32_e32 v16, v19, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v20, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v16, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v15
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v14, v15, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v11, v13, s4
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v13
; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s4
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v12, v12, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v9, v11, s4
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v11
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v10, v11, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v7, v9, s4
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v9
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s4
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v9, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v7, s4
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s4
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v7, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v3, v5, s4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v5
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v4, v5, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    v_perm_b32 v1, v2, v41, s4
; GFX9-NEXT:    v_perm_b32 v2, v40, v42, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v128i8_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:72
; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:76
; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v74, s30, 0
; GFX11-NEXT:    v_writelane_b32 v75, s96, 0
; GFX11-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_writelane_b32 v74, s31, 1
; GFX11-NEXT:    v_writelane_b32 v75, s97, 1
; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
; GFX11-NEXT:    v_writelane_b32 v74, s34, 2
; GFX11-NEXT:    v_writelane_b32 v75, s98, 2
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_writelane_b32 v74, s35, 3
; GFX11-NEXT:    v_writelane_b32 v75, s99, 3
; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
; GFX11-NEXT:    v_writelane_b32 v74, s36, 4
; GFX11-NEXT:    v_writelane_b32 v75, s100, 4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_writelane_b32 v74, s37, 5
; GFX11-NEXT:    v_writelane_b32 v75, s101, 5
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:4
; GFX11-NEXT:    scratch_store_b32 off, v73, s32
; GFX11-NEXT:    v_writelane_b32 v74, s38, 6
; GFX11-NEXT:    v_writelane_b32 v75, s102, 6
; GFX11-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
; GFX11-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
; GFX11-NEXT:    v_writelane_b32 v74, s39, 7
; GFX11-NEXT:    v_writelane_b32 v75, s103, 7
; GFX11-NEXT:    v_writelane_b32 v74, s48, 8
; GFX11-NEXT:    v_writelane_b32 v75, s104, 8
; GFX11-NEXT:    v_writelane_b32 v74, s49, 9
; GFX11-NEXT:    v_writelane_b32 v74, s50, 10
; GFX11-NEXT:    v_writelane_b32 v74, s51, 11
; GFX11-NEXT:    v_writelane_b32 v74, s52, 12
; GFX11-NEXT:    v_writelane_b32 v74, s53, 13
; GFX11-NEXT:    v_writelane_b32 v74, s54, 14
; GFX11-NEXT:    v_writelane_b32 v74, s55, 15
; GFX11-NEXT:    v_writelane_b32 v74, s64, 16
; GFX11-NEXT:    v_writelane_b32 v74, s65, 17
; GFX11-NEXT:    v_writelane_b32 v74, s66, 18
; GFX11-NEXT:    v_writelane_b32 v74, s67, 19
; GFX11-NEXT:    v_writelane_b32 v74, s68, 20
; GFX11-NEXT:    v_writelane_b32 v74, s69, 21
; GFX11-NEXT:    v_writelane_b32 v74, s70, 22
; GFX11-NEXT:    v_writelane_b32 v74, s71, 23
; GFX11-NEXT:    v_writelane_b32 v74, s80, 24
; GFX11-NEXT:    v_writelane_b32 v74, s81, 25
; GFX11-NEXT:    v_writelane_b32 v74, s82, 26
; GFX11-NEXT:    v_writelane_b32 v74, s83, 27
; GFX11-NEXT:    v_writelane_b32 v74, s84, 28
; GFX11-NEXT:    v_writelane_b32 v74, s85, 29
; GFX11-NEXT:    v_writelane_b32 v74, s86, 30
; GFX11-NEXT:    v_writelane_b32 v74, s87, 31
; GFX11-NEXT:    s_cbranch_scc0 .LBB95_3
; GFX11-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s5, 24
; GFX11-NEXT:    v_writelane_b32 v77, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
; GFX11-NEXT:    s_lshr_b32 s34, s5, 16
; GFX11-NEXT:    s_lshr_b32 s36, s5, 8
; GFX11-NEXT:    s_lshr_b32 s35, s4, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s27, 8
; GFX11-NEXT:    s_lshr_b32 s37, s4, 8
; GFX11-NEXT:    s_lshr_b32 s38, s7, 24
; GFX11-NEXT:    s_lshr_b32 s39, s7, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s26, 16
; GFX11-NEXT:    s_lshr_b32 s49, s7, 8
; GFX11-NEXT:    s_lshr_b32 s48, s6, 16
; GFX11-NEXT:    s_lshr_b32 s50, s6, 8
; GFX11-NEXT:    v_writelane_b32 v77, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s26, 8
; GFX11-NEXT:    s_lshr_b32 s51, s9, 24
; GFX11-NEXT:    s_lshr_b32 s52, s9, 16
; GFX11-NEXT:    s_lshr_b32 s54, s9, 8
; GFX11-NEXT:    v_writelane_b32 v77, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s25, 24
; GFX11-NEXT:    s_lshr_b32 s53, s8, 16
; GFX11-NEXT:    s_lshr_b32 s55, s8, 8
; GFX11-NEXT:    s_lshr_b32 s64, s11, 24
; GFX11-NEXT:    v_writelane_b32 v77, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
; GFX11-NEXT:    s_lshr_b32 s65, s11, 16
; GFX11-NEXT:    s_lshr_b32 s67, s11, 8
; GFX11-NEXT:    s_lshr_b32 s66, s10, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
; GFX11-NEXT:    s_lshr_b32 s68, s10, 8
; GFX11-NEXT:    s_lshr_b32 s69, s13, 24
; GFX11-NEXT:    s_lshr_b32 s70, s13, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s24, 16
; GFX11-NEXT:    s_lshr_b32 s80, s13, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s24, 8
; GFX11-NEXT:    s_lshr_b32 s71, s12, 16
; GFX11-NEXT:    s_lshr_b32 s81, s12, 8
; GFX11-NEXT:    s_lshr_b32 s82, s15, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
; GFX11-NEXT:    s_lshr_b32 s83, s15, 16
; GFX11-NEXT:    s_lshr_b32 s85, s15, 8
; GFX11-NEXT:    s_lshr_b32 s84, s14, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
; GFX11-NEXT:    s_lshr_b32 s86, s14, 8
; GFX11-NEXT:    s_lshr_b32 s87, s41, 24
; GFX11-NEXT:    s_lshr_b32 s96, s41, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
; GFX11-NEXT:    s_lshr_b32 s98, s41, 8
; GFX11-NEXT:    s_lshr_b32 s97, s40, 16
; GFX11-NEXT:    s_lshr_b32 s99, s40, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s22, 16
; GFX11-NEXT:    s_lshr_b32 s100, s29, 24
; GFX11-NEXT:    s_lshr_b32 s101, s29, 16
; GFX11-NEXT:    s_lshr_b32 s103, s29, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s22, 8
; GFX11-NEXT:    s_lshr_b32 s102, s28, 16
; GFX11-NEXT:    s_lshr_b32 s104, s28, 8
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s20, 8
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s18, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s18, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s16, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s2, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s2, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s1, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s0, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s0, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 0
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-NEXT:    s_cbranch_vccnz .LBB95_4
; GFX11-NEXT:  .LBB95_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, s18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, s17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s16 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s41 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s40 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s9 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s8 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v34, 0x200, s3 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v33, 0x200, s2 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, s23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s29 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s28 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s11 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s10 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s7 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s6 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s5 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s4 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v36, 0x200, s1 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v35, 0x200, s0 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_lshrrev_b64 v[51:52], 24, v[17:18]
; GFX11-NEXT:    v_lshrrev_b64 v[64:65], 24, v[27:28]
; GFX11-NEXT:    v_lshrrev_b64 v[52:53], 24, v[19:20]
; GFX11-NEXT:    v_lshrrev_b64 v[65:66], 24, v[29:30]
; GFX11-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX11-NEXT:    v_lshrrev_b64 v[69:70], 24, v[33:34]
; GFX11-NEXT:    v_lshrrev_b64 v[37:38], 24, v[5:6]
; GFX11-NEXT:    v_lshrrev_b64 v[48:49], 24, v[9:10]
; GFX11-NEXT:    v_lshrrev_b64 v[66:67], 24, v[13:14]
; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-NEXT:    v_lshrrev_b64 v[70:71], 24, v[35:36]
; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[1:2]
; GFX11-NEXT:    v_lshrrev_b64 v[31:32], 24, v[3:4]
; GFX11-NEXT:    v_lshrrev_b64 v[38:39], 24, v[7:8]
; GFX11-NEXT:    v_lshrrev_b64 v[49:50], 24, v[11:12]
; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[15:16]
; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 24, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 16, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 8, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 16, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 8, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 24, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 16, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 16, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 8, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 24, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 16, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 8, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 16, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 8, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 24, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 16, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 8, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 24, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 16, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 8, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 16, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 8, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 24, v30
; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 16, v30
; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 8, v30
; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 16, v29
; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 8, v29
; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 24, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 16, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 8, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 16, v33
; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 8, v33
; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 24, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 16, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 8, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 16, v35
; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 8, v35
; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 24, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 8, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 8, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 24, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 8, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 16, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 8, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 24, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 8, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 16, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 8, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 24, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 16, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 8, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 8, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 24, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 16, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 8, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 16, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 8, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 24, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 16, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 8, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 8, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 24, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 16, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 16, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 8, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 24, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 16, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 8, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 16, v15
; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 8, v15
; GFX11-NEXT:    s_branch .LBB95_5
; GFX11-NEXT:  .LBB95_3:
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr44
; GFX11-NEXT:    ; implicit-def: $sgpr46
; GFX11-NEXT:    ; implicit-def: $sgpr56
; GFX11-NEXT:    ; implicit-def: $sgpr58
; GFX11-NEXT:    ; implicit-def: $sgpr60
; GFX11-NEXT:    ; implicit-def: $sgpr104
; GFX11-NEXT:    ; implicit-def: $sgpr102
; GFX11-NEXT:    ; implicit-def: $sgpr103
; GFX11-NEXT:    ; implicit-def: $sgpr101
; GFX11-NEXT:    ; implicit-def: $sgpr100
; GFX11-NEXT:    ; implicit-def: $sgpr99
; GFX11-NEXT:    ; implicit-def: $sgpr97
; GFX11-NEXT:    ; implicit-def: $sgpr98
; GFX11-NEXT:    ; implicit-def: $sgpr96
; GFX11-NEXT:    ; implicit-def: $sgpr87
; GFX11-NEXT:    ; implicit-def: $sgpr86
; GFX11-NEXT:    ; implicit-def: $sgpr84
; GFX11-NEXT:    ; implicit-def: $sgpr85
; GFX11-NEXT:    ; implicit-def: $sgpr83
; GFX11-NEXT:    ; implicit-def: $sgpr82
; GFX11-NEXT:    ; implicit-def: $sgpr81
; GFX11-NEXT:    ; implicit-def: $sgpr71
; GFX11-NEXT:    ; implicit-def: $sgpr80
; GFX11-NEXT:    ; implicit-def: $sgpr70
; GFX11-NEXT:    ; implicit-def: $sgpr69
; GFX11-NEXT:    ; implicit-def: $sgpr68
; GFX11-NEXT:    ; implicit-def: $sgpr66
; GFX11-NEXT:    ; implicit-def: $sgpr67
; GFX11-NEXT:    ; implicit-def: $sgpr65
; GFX11-NEXT:    ; implicit-def: $sgpr64
; GFX11-NEXT:    ; implicit-def: $sgpr55
; GFX11-NEXT:    ; implicit-def: $sgpr53
; GFX11-NEXT:    ; implicit-def: $sgpr54
; GFX11-NEXT:    ; implicit-def: $sgpr52
; GFX11-NEXT:    ; implicit-def: $sgpr51
; GFX11-NEXT:    ; implicit-def: $sgpr50
; GFX11-NEXT:    ; implicit-def: $sgpr48
; GFX11-NEXT:    ; implicit-def: $sgpr49
; GFX11-NEXT:    ; implicit-def: $sgpr39
; GFX11-NEXT:    ; implicit-def: $sgpr38
; GFX11-NEXT:    ; implicit-def: $sgpr37
; GFX11-NEXT:    ; implicit-def: $sgpr35
; GFX11-NEXT:    ; implicit-def: $sgpr36
; GFX11-NEXT:    ; implicit-def: $sgpr34
; GFX11-NEXT:    ; implicit-def: $vcc_hi
; GFX11-NEXT:    ; implicit-def: $sgpr72
; GFX11-NEXT:    ; implicit-def: $sgpr62
; GFX11-NEXT:    ; implicit-def: $sgpr30
; GFX11-NEXT:    ; implicit-def: $sgpr94
; GFX11-NEXT:    ; implicit-def: $sgpr92
; GFX11-NEXT:    ; implicit-def: $sgpr90
; GFX11-NEXT:    ; implicit-def: $sgpr88
; GFX11-NEXT:    ; implicit-def: $sgpr78
; GFX11-NEXT:    ; implicit-def: $sgpr76
; GFX11-NEXT:    ; implicit-def: $sgpr74
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    s_branch .LBB95_2
; GFX11-NEXT:  .LBB95_4:
; GFX11-NEXT:    v_dual_mov_b32 v35, s0 :: v_dual_mov_b32 v36, s1
; GFX11-NEXT:    v_readlane_b32 s0, v76, 0
; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v63, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 1
; GFX11-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
; GFX11-NEXT:    v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
; GFX11-NEXT:    v_mov_b32_e32 v73, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 2
; GFX11-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
; GFX11-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v61, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 3
; GFX11-NEXT:    v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
; GFX11-NEXT:    v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
; GFX11-NEXT:    v_mov_b32_e32 v62, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 4
; GFX11-NEXT:    v_dual_mov_b32 v33, s2 :: v_dual_mov_b32 v34, s3
; GFX11-NEXT:    v_dual_mov_b32 v29, s16 :: v_dual_mov_b32 v30, s17
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v72, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 5
; GFX11-NEXT:    v_dual_mov_b32 v27, s18 :: v_dual_mov_b32 v28, s19
; GFX11-NEXT:    v_dual_mov_b32 v23, s20 :: v_dual_mov_b32 v24, s21
; GFX11-NEXT:    v_mov_b32_e32 v56, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 6
; GFX11-NEXT:    v_dual_mov_b32 v21, s22 :: v_dual_mov_b32 v22, s23
; GFX11-NEXT:    v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v60, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 7
; GFX11-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
; GFX11-NEXT:    v_dual_mov_b32 v134, s104 :: v_dual_mov_b32 v145, s102
; GFX11-NEXT:    v_mov_b32_e32 v57, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 8
; GFX11-NEXT:    v_dual_mov_b32 v132, s103 :: v_dual_mov_b32 v135, s101
; GFX11-NEXT:    v_dual_mov_b32 v144, s100 :: v_dual_mov_b32 v131, s99
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v58, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 9
; GFX11-NEXT:    v_dual_mov_b32 v133, s97 :: v_dual_mov_b32 v128, s98
; GFX11-NEXT:    v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v130, s87
; GFX11-NEXT:    v_mov_b32_e32 v59, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 10
; GFX11-NEXT:    v_dual_mov_b32 v116, s86 :: v_dual_mov_b32 v119, s84
; GFX11-NEXT:    v_dual_mov_b32 v114, s85 :: v_dual_mov_b32 v117, s83
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v45, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 11
; GFX11-NEXT:    v_dual_mov_b32 v118, s82 :: v_dual_mov_b32 v113, s81
; GFX11-NEXT:    v_dual_mov_b32 v115, s71 :: v_dual_mov_b32 v102, s80
; GFX11-NEXT:    v_mov_b32_e32 v47, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 12
; GFX11-NEXT:    v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v112, s69
; GFX11-NEXT:    v_dual_mov_b32 v98, s68 :: v_dual_mov_b32 v101, s66
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v40, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 13
; GFX11-NEXT:    v_dual_mov_b32 v96, s67 :: v_dual_mov_b32 v99, s65
; GFX11-NEXT:    v_dual_mov_b32 v100, s64 :: v_dual_mov_b32 v87, s55
; GFX11-NEXT:    v_mov_b32_e32 v44, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 14
; GFX11-NEXT:    v_dual_mov_b32 v97, s53 :: v_dual_mov_b32 v86, s51
; GFX11-NEXT:    v_dual_mov_b32 v83, s54 :: v_dual_mov_b32 v82, s50
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v46, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 15
; GFX11-NEXT:    v_dual_mov_b32 v85, s52 :: v_dual_mov_b32 v84, s48
; GFX11-NEXT:    v_dual_mov_b32 v55, s49 :: v_dual_mov_b32 v80, s39
; GFX11-NEXT:    v_mov_b32_e32 v182, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 16
; GFX11-NEXT:    v_dual_mov_b32 v81, s38 :: v_dual_mov_b32 v32, s37
; GFX11-NEXT:    v_dual_mov_b32 v71, s35 :: v_dual_mov_b32 v26, s36
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v43, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 17
; GFX11-NEXT:    v_dual_mov_b32 v39, s34 :: v_dual_mov_b32 v50, vcc_hi
; GFX11-NEXT:    v_dual_mov_b32 v51, s62 :: v_dual_mov_b32 v52, s72
; GFX11-NEXT:    v_mov_b32_e32 v183, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 18
; GFX11-NEXT:    v_dual_mov_b32 v53, s60 :: v_dual_mov_b32 v54, s58
; GFX11-NEXT:    v_dual_mov_b32 v64, s56 :: v_dual_mov_b32 v65, s46
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v41, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 19
; GFX11-NEXT:    v_dual_mov_b32 v69, s44 :: v_dual_mov_b32 v70, s42
; GFX11-NEXT:    v_dual_mov_b32 v25, s74 :: v_dual_mov_b32 v38, s88
; GFX11-NEXT:    v_mov_b32_e32 v42, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 20
; GFX11-NEXT:    v_dual_mov_b32 v31, s76 :: v_dual_mov_b32 v48, s90
; GFX11-NEXT:    v_dual_mov_b32 v37, s78 :: v_dual_mov_b32 v66, s94
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v177, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 21
; GFX11-NEXT:    v_mov_b32_e32 v49, s92
; GFX11-NEXT:    v_mov_b32_e32 v67, s30
; GFX11-NEXT:    v_mov_b32_e32 v181, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 22
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v167, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 23
; GFX11-NEXT:    v_mov_b32_e32 v179, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 24
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v180, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 25
; GFX11-NEXT:    v_mov_b32_e32 v176, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 26
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v178, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 27
; GFX11-NEXT:    v_mov_b32_e32 v164, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 28
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v165, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 29
; GFX11-NEXT:    v_mov_b32_e32 v166, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 30
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v151, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 31
; GFX11-NEXT:    v_mov_b32_e32 v163, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v149, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 1
; GFX11-NEXT:    v_mov_b32_e32 v161, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v162, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 3
; GFX11-NEXT:    v_mov_b32_e32 v150, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v160, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 5
; GFX11-NEXT:    v_mov_b32_e32 v146, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v147, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 7
; GFX11-NEXT:    v_mov_b32_e32 v148, s0
; GFX11-NEXT:  .LBB95_5: ; %end
; GFX11-NEXT:    v_perm_b32 v68, v73, v70, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v70, v62, v72, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v69, v60, v69, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v36, v36, v61, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v65, v47, v65, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v64, v43, v64, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 16, v70
; GFX11-NEXT:    v_perm_b32 v35, v35, v63, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 16, v68
; GFX11-NEXT:    v_perm_b32 v56, v33, v56, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
; GFX11-NEXT:    v_perm_b32 v58, v58, v59, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v57, v34, v57, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v34, v36, v70
; GFX11-NEXT:    v_perm_b32 v29, v29, v45, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX11-NEXT:    v_perm_b32 v70, v27, v182, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX11-NEXT:    v_or_b32_e32 v33, v35, v68
; GFX11-NEXT:    v_or_b32_e32 v35, v56, v69
; GFX11-NEXT:    v_perm_b32 v68, v44, v46, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v69, v41, v42, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v58, 16, v58
; GFX11-NEXT:    v_or_b32_e32 v27, v29, v65
; GFX11-NEXT:    v_or_b32_e32 v29, v70, v64
; GFX11-NEXT:    v_perm_b32 v54, v181, v54, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v64, v179, v180, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v53, v178, v53, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v30, v30, v40, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 16, v68
; GFX11-NEXT:    v_perm_b32 v182, v28, v183, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
; GFX11-NEXT:    v_or_b32_e32 v36, v57, v58
; GFX11-NEXT:    v_perm_b32 v23, v23, v177, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX11-NEXT:    v_perm_b32 v24, v24, v167, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX11-NEXT:    v_perm_b32 v21, v21, v176, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX11-NEXT:    v_or_b32_e32 v28, v30, v68
; GFX11-NEXT:    v_or_b32_e32 v30, v182, v69
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_store_b128 v0, v[33:36], off
; GFX11-NEXT:    scratch_store_b128 v0, v[27:30], off offset:16
; GFX11-NEXT:    v_or_b32_e32 v27, v23, v54
; GFX11-NEXT:    v_perm_b32 v23, v165, v166, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v28, v24, v64
; GFX11-NEXT:    v_or_b32_e32 v29, v21, v53
; GFX11-NEXT:    v_perm_b32 v21, v22, v164, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v22, v163, v52, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v161, v162, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v30, v160, v51, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v17, v17, v150, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v30
; GFX11-NEXT:    v_or_b32_e32 v30, v21, v23
; GFX11-NEXT:    v_or_b32_e32 v19, v19, v22
; GFX11-NEXT:    v_perm_b32 v22, v147, v148, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v20, v20, v24
; GFX11-NEXT:    v_or_b32_e32 v21, v17, v33
; GFX11-NEXT:    v_perm_b32 v17, v18, v146, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v18, v145, v67, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v133, v66, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v135, v144, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v15, v15, v134, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v13, v13, v131, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v16, v16, v132, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_or_b32_e32 v22, v17, v22
; GFX11-NEXT:    v_or_b32_e32 v15, v15, v18
; GFX11-NEXT:    v_perm_b32 v18, v129, v130, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v17, v13, v24
; GFX11-NEXT:    v_perm_b32 v24, v115, v48, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v16, v16, v23
; GFX11-NEXT:    v_perm_b32 v13, v14, v128, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v14, v119, v49, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v117, v118, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v9, v9, v113, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v11, v11, v116, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v12, v12, v114, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_or_b32_e32 v18, v13, v18
; GFX11-NEXT:    v_or_b32_e32 v13, v9, v24
; GFX11-NEXT:    v_perm_b32 v9, v10, v102, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v10, v101, v38, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v11, v11, v14
; GFX11-NEXT:    v_perm_b32 v14, v103, v112, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v12, v12, v23
; GFX11-NEXT:    v_perm_b32 v23, v99, v100, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v97, v37, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v7, v7, v98, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v8, v8, v96, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v5, v5, v87, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_or_b32_e32 v7, v7, v10
; GFX11-NEXT:    v_perm_b32 v10, v85, v86, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v14, v9, v14
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v23
; GFX11-NEXT:    v_or_b32_e32 v9, v5, v24
; GFX11-NEXT:    v_perm_b32 v5, v84, v31, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v80, v81, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v71, v25, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v25, v39, v50, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v6, v6, v83, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_perm_b32 v3, v3, v82, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-NEXT:    v_perm_b32 v4, v4, v55, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v31, v1, v32, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v26, v2, v26, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-NEXT:    v_or_b32_e32 v10, v6, v10
; GFX11-NEXT:    v_or_b32_e32 v1, v3, v5
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v23
; GFX11-NEXT:    v_or_b32_e32 v3, v31, v24
; GFX11-NEXT:    v_or_b32_e32 v4, v26, v25
; GFX11-NEXT:    s_clause 0x5
; GFX11-NEXT:    scratch_store_b128 v0, v[27:30], off offset:32
; GFX11-NEXT:    scratch_store_b128 v0, v[19:22], off offset:48
; GFX11-NEXT:    scratch_store_b128 v0, v[15:18], off offset:64
; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v73, off, s32
; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-NEXT:    v_readlane_b32 s104, v75, 8
; GFX11-NEXT:    v_readlane_b32 s103, v75, 7
; GFX11-NEXT:    v_readlane_b32 s102, v75, 6
; GFX11-NEXT:    v_readlane_b32 s101, v75, 5
; GFX11-NEXT:    v_readlane_b32 s100, v75, 4
; GFX11-NEXT:    v_readlane_b32 s99, v75, 3
; GFX11-NEXT:    v_readlane_b32 s98, v75, 2
; GFX11-NEXT:    v_readlane_b32 s97, v75, 1
; GFX11-NEXT:    v_readlane_b32 s96, v75, 0
; GFX11-NEXT:    v_readlane_b32 s87, v74, 31
; GFX11-NEXT:    v_readlane_b32 s86, v74, 30
; GFX11-NEXT:    v_readlane_b32 s85, v74, 29
; GFX11-NEXT:    v_readlane_b32 s84, v74, 28
; GFX11-NEXT:    v_readlane_b32 s83, v74, 27
; GFX11-NEXT:    v_readlane_b32 s82, v74, 26
; GFX11-NEXT:    v_readlane_b32 s81, v74, 25
; GFX11-NEXT:    v_readlane_b32 s80, v74, 24
; GFX11-NEXT:    v_readlane_b32 s71, v74, 23
; GFX11-NEXT:    v_readlane_b32 s70, v74, 22
; GFX11-NEXT:    v_readlane_b32 s69, v74, 21
; GFX11-NEXT:    v_readlane_b32 s68, v74, 20
; GFX11-NEXT:    v_readlane_b32 s67, v74, 19
; GFX11-NEXT:    v_readlane_b32 s66, v74, 18
; GFX11-NEXT:    v_readlane_b32 s65, v74, 17
; GFX11-NEXT:    v_readlane_b32 s64, v74, 16
; GFX11-NEXT:    v_readlane_b32 s55, v74, 15
; GFX11-NEXT:    v_readlane_b32 s54, v74, 14
; GFX11-NEXT:    v_readlane_b32 s53, v74, 13
; GFX11-NEXT:    v_readlane_b32 s52, v74, 12
; GFX11-NEXT:    v_readlane_b32 s51, v74, 11
; GFX11-NEXT:    v_readlane_b32 s50, v74, 10
; GFX11-NEXT:    v_readlane_b32 s49, v74, 9
; GFX11-NEXT:    v_readlane_b32 s48, v74, 8
; GFX11-NEXT:    v_readlane_b32 s39, v74, 7
; GFX11-NEXT:    v_readlane_b32 s38, v74, 6
; GFX11-NEXT:    v_readlane_b32 s37, v74, 5
; GFX11-NEXT:    v_readlane_b32 s36, v74, 4
; GFX11-NEXT:    v_readlane_b32 s35, v74, 3
; GFX11-NEXT:    v_readlane_b32 s34, v74, 2
; GFX11-NEXT:    v_readlane_b32 s31, v74, 1
; GFX11-NEXT:    v_readlane_b32 s30, v74, 0
; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:72
; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:76
; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define <64 x i16> @bitcast_v128i8_to_v64i16(<128 x i8> %a, i32 %b) {
; SI-LABEL: bitcast_v128i8_to_v64i16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v59, v24
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v47, v16
; SI-NEXT:    v_mov_b32_e32 v61, v14
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v60, v6
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:388
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:48
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:80
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:92
; SI-NEXT:    v_lshlrev_b32_e32 v40, 8, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v25
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:4
; SI-NEXT:    v_lshlrev_b32_e32 v5, 8, v5
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 24, v7
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v13
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v15
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v11
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v21
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v23
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v19
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 8, v29
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v27
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v43, v8
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v26
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v24
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:124
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:116
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:76
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:36
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:108
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:28
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v6
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:100
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:136
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:132
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(5) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v5
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v30
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v31
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:176
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:168
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v10
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v32
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v33
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v34
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:188
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:196
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 24, v2
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v3
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:220
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v11
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:872 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:204
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:240
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:236
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:252
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:844 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v2
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:232
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:228
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:888 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v12
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:920 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:896 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:268
; SI-NEXT:    v_lshlrev_b32_e32 v63, 8, v2
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:852 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:916 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:316
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:264
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:260
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 24, v0
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:876 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:340
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:300
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v13
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:952 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:928 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:336
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:332
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:860 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:296
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:292
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v56, 24, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:936 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:328
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:324
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 24, v3
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v14
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:956 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 8, v2
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:352
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:376
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v46, 24, v0
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:384
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:380
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:948 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:372
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:368
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:364
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:360
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:356
; SI-NEXT:    v_lshlrev_b32_e32 v58, 8, v18
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v57, 24, v15
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v62, 24, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 8, v20
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:68
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:904 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:60
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:20
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:44
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB96_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v51, 0xff, v47
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v48, 0xff, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_or_b32_e32 v0, v57, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v58
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v62, v3
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v5, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v60
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_mov_b32_e32 v60, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v5, v1, v5
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v9, v1, v9
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v10, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v61
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v10, v1, v10
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v1, v12
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v13, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v1, v14
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v1, v14
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v15, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v1, v16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v1, v16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v19, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v49, v1, v20
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v20, v1, v20
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v21, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v50, v1, v22
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v22, v1, v22
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v23, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v52, v1, v24
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v24, v1, v24
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v27, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v53, v1, v28
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v28, v1, v28
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v29, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v54, v1, v30
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v1, v30
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v33, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v55, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v34, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v35, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v41, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v36, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v37, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v42, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v38, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v39, v31, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v44, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v45, v1, v31
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v63, v31, v63
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v26, v26, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v7, v7, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v32, v31, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v46, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v46, v56, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v31, v31, v40
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v59
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v31, v9
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v48, v48, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v51, v51, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v56, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v57, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v58, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v62, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v6, v40, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v40, v1
; SI-NEXT:    v_and_b32_e32 v40, 0xff, v60
; SI-NEXT:    v_or_b32_e32 v40, v40, v43
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v60, v43, v47
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v47, v43, v47
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    v_or_b32_e32 v61, v43, v59
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v43, 0xff, v43
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v43, v43, v59
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v59, 0xff, v59
; SI-NEXT:    v_or_b32_e32 v59, v59, v8
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v5
; SI-NEXT:    v_alignbit_b32 v9, v8, v9, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:968 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v11, v9, v10
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v48
; SI-NEXT:    v_or_b32_e32 v48, v9, v12
; SI-NEXT:    v_alignbit_b32 v9, v11, v12, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v12, v9, v17
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v51
; SI-NEXT:    v_or_b32_e32 v9, v9, v14
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v9, v12, v14, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v51, v9, v25
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v56
; SI-NEXT:    v_or_b32_e32 v15, v9, v16
; SI-NEXT:    v_alignbit_b32 v9, v51, v16, 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v9, v9, v49
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v57
; SI-NEXT:    v_or_b32_e32 v16, v13, v20
; SI-NEXT:    v_alignbit_b32 v13, v9, v20, 16
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v31, v13, v50
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v58
; SI-NEXT:    v_or_b32_e32 v13, v13, v22
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v13, v31, v22, 16
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v62
; SI-NEXT:    v_or_b32_e32 v13, v13, v52
; SI-NEXT:    v_or_b32_e32 v14, v14, v24
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v14, v13, v24, 16
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v24, v14, v53
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v14, v28
; SI-NEXT:    v_alignbit_b32 v14, v24, v28, 16
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v21, v14, v54
; SI-NEXT:    v_or_b32_e32 v22, v6, v30
; SI-NEXT:    v_alignbit_b32 v6, v21, v30, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v33
; SI-NEXT:    v_or_b32_e32 v19, v6, v55
; SI-NEXT:    v_or_b32_e32 v29, v1, v34
; SI-NEXT:    v_alignbit_b32 v1, v19, v34, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v35
; SI-NEXT:    v_or_b32_e32 v23, v1, v41
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v40
; SI-NEXT:    v_or_b32_e32 v20, v1, v36
; SI-NEXT:    v_alignbit_b32 v1, v23, v36, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v37
; SI-NEXT:    v_or_b32_e32 v30, v1, v42
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v60
; SI-NEXT:    v_or_b32_e32 v27, v1, v38
; SI-NEXT:    v_alignbit_b32 v1, v30, v38, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v39
; SI-NEXT:    v_or_b32_e32 v28, v1, v44
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v47
; SI-NEXT:    v_or_b32_e32 v34, v1, v45
; SI-NEXT:    v_alignbit_b32 v1, v28, v45, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v63
; SI-NEXT:    v_or_b32_e32 v33, v1, v26
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
; SI-NEXT:    v_or_b32_e32 v36, v1, v7
; SI-NEXT:    v_alignbit_b32 v1, v33, v7, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v32
; SI-NEXT:    v_or_b32_e32 v35, v1, v4
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v43
; SI-NEXT:    v_or_b32_e32 v38, v1, v46
; SI-NEXT:    v_alignbit_b32 v1, v35, v46, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v37, v1, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v59
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_or_b32_e32 v39, v1, v3
; SI-NEXT:    v_alignbit_b32 v1, v37, v3, 16
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v50
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v52
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v54
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v41
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v42
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v44
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v4
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:  .LBB96_2: ; %Flow
; SI-NEXT:    s_or_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:968 ; 4-byte Folded Reload
; SI-NEXT:    s_xor_b64 exec, exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB96_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:904 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_or_b32_e32 v2, v58, v2
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_movk_i32 s6, 0x300
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_or_b32_e32 v0, v57, v0
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v3, v62, v3
; SI-NEXT:    s_mov_b32 s7, 0x3000000
; SI-NEXT:    v_add_i32_e32 v37, vcc, s7, v0
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v37
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:916 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:876 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:896 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:844 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:964 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_or_b32_e32 v6, v26, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v5, v1, v5
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, 0x300, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v3, v3, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:956 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v39, vcc, s7, v3
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:960 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:936 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v56, v2
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:928 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v38, vcc, s7, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:948 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v4, v46, v4
; SI-NEXT:    v_or_b32_e32 v2, v4, v2
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v35, vcc, s7, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v35
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xff, v4
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:860 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_or_b32_e32 v5, v7, v5
; SI-NEXT:    v_or_b32_e32 v4, v5, v4
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:852 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:952 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v36, vcc, s7, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xff, v5
; SI-NEXT:    v_or_b32_e32 v5, v63, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v33, vcc, s7, v5
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v33
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xff, v6
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_or_b32_e32 v6, v7, v6
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:888 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v34, vcc, s7, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xff, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:920 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v28, vcc, s7, v7
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xff, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_or_b32_e32 v8, v9, v8
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v27, vcc, s7, v8
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xff, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:872 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v32
; SI-NEXT:    v_and_b32_e32 v10, 0xff, v10
; SI-NEXT:    v_add_i32_e32 v30, vcc, s7, v9
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_or_b32_e32 v10, v11, v10
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xff, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v23, vcc, s7, v11
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xff, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_or_b32_e32 v12, v13, v12
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s7, v12
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xff, v13
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xff, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_or_b32_e32 v14, v15, v14
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v22, vcc, s7, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    v_or_b32_e32 v16, v17, v16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v5, vcc, s7, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v24, vcc, s7, v17
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_or_b32_e32 v18, v19, v18
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v4, vcc, s7, v18
; SI-NEXT:    v_mov_b32_e32 v18, v5
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v25, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v26, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v3, vcc, s7, v26
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_or_b32_e32 v20, v21, v20
; SI-NEXT:    v_or_b32_e32 v31, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v21, vcc, s7, v15
; SI-NEXT:    v_add_i32_e32 v31, vcc, s7, v31
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v32, v20, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v32
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v48, v20, v19
; SI-NEXT:    v_or_b32_e32 v0, v48, v0
; SI-NEXT:    v_add_i32_e32 v16, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v19, vcc, s7, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, s7, v25
; SI-NEXT:    v_add_i32_e32 v20, vcc, s7, v10
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v9, vcc, s7, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v59
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v15, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v51, vcc, s7, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v47
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v51
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v6, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v12, vcc, s7, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v43
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v48, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v61
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v11, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_or_b32_e32 v0, v40, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v7, vcc, s7, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xff, v0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v60
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v8, vcc, s7, v0
; SI-NEXT:    v_alignbit_b32 v0, v11, v48, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:840 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v12, v6, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:848 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v51, v15, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:856 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v9, v16, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:864 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v31, v3, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:868 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v13, v4, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:880 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v24, v5, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:884 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v21, v22, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:892 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v19, v29, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:900 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v23, v20, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:908 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v30, v27, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:912 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v28, v34, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:924 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v33, v36, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:932 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v35, v38, 16
; SI-NEXT:    v_alignbit_b32 v1, v8, v7, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:940 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v0, v37, v39, 16
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:944 ; 4-byte Folded Spill
; SI-NEXT:  .LBB96_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:840 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:856 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v6
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v41
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v45
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:848 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:880 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v48
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:864 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:868 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v51
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v8, v8, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v17
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:884 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v49
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:892 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v52
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:944 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v50
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:900 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v25
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v29
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:924 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v54
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v55
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:940 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v18, v18, v21
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v53
; SI-NEXT:    v_or_b32_e32 v19, v19, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:908 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:912 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:932 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v34
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v44
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v36
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v33
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v38
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v35
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v39
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v37
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v128i8_to_v64i16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:112
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:64
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:40
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:32
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:24
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:16
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:12
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:4
; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:384
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:288
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:280
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:272
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:268
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:380
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:376
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:372
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:364
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:360
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:356
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:352
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:348
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:344
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:340
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:332
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:160
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:328
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:156
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:324
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:152
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:148
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:316
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:144
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:312
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:140
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:136
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:304
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:132
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:300
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:120
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB96_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s6, 0xc0c0004
; VI-NEXT:    v_perm_b32 v10, v37, v36, s6
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v11, v59, v34, s6
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; VI-NEXT:    v_perm_b32 v12, v57, v32, s6
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_perm_b32 v13, v47, v46, s6
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_perm_b32 v14, v43, v42, s6
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_perm_b32 v15, v55, v54, s6
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v16, v51, v50, s6
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:    ; implicit-def: $vgpr36
; VI-NEXT:    ; implicit-def: $vgpr59
; VI-NEXT:    ; implicit-def: $vgpr34
; VI-NEXT:    ; implicit-def: $vgpr57
; VI-NEXT:    ; implicit-def: $vgpr47
; VI-NEXT:    ; implicit-def: $vgpr46
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr55
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    ; implicit-def: $vgpr51
; VI-NEXT:    ; implicit-def: $vgpr50
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v1, v0, s6
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s6
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s6
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v3, v4, v3, s6
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v4, v5, v4, s6
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v5, v6, v5, s6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v6, v7, v6, s6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v7, v8, v7, s6
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v8, v8, v9, s6
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_perm_b32 v9, v39, v62, s6
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_perm_b32 v8, v38, v61, s6
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_perm_b32 v9, v35, v60, s6
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_perm_b32 v10, v33, v58, s6
; VI-NEXT:    v_or_b32_e32 v10, v10, v11
; VI-NEXT:    v_perm_b32 v11, v63, v56, s6
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    v_perm_b32 v12, v45, v44, s6
; VI-NEXT:    v_or_b32_e32 v12, v12, v13
; VI-NEXT:    v_perm_b32 v13, v41, v40, s6
; VI-NEXT:    v_or_b32_e32 v13, v13, v14
; VI-NEXT:    v_perm_b32 v14, v53, v52, s6
; VI-NEXT:    v_or_b32_e32 v14, v14, v15
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr61
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr62
; VI-NEXT:    ; implicit-def: $vgpr35
; VI-NEXT:    ; implicit-def: $vgpr60
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr63
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr45
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr53
; VI-NEXT:    ; implicit-def: $vgpr52
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v49, v15, s6
; VI-NEXT:    v_or_b32_e32 v15, v15, v16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v17, v16, s6
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v18, v17, s6
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v17, v18, v19
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v18, v19, v18, s6
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v18, v18, v19
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v20, v19, s6
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v21, v20, s6
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v19, v19, v20
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v20, v20, v21, s6
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v22, v21, s6
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v20, v20, v21
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v21, v21, v22, s6
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v21, v21, v22
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v22, v23, v22, s6
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v23, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v22, v22, v23
; VI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v23, v24, v23, s6
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_or_b32_e32 v23, v23, v24
; VI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v24, v25, v24, s6
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v25, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_e32 v24, v24, v25
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v25, v26, v25, s6
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    v_or_b32_e32 v25, v25, v26
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s6
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v28, v27, s6
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    v_or_b32_e32 v26, v26, v27
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v27, v28, s6
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    v_or_b32_e32 v27, v27, v28
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v29, v28, s6
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v30, v29, s6
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v28, v28, v29
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v29, v30, s6
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    v_or_b32_e32 v29, v29, v30
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s6
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v32, v31, s6
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_or_b32_e32 v30, v30, v31
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v31, v31, v48, s6
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v32, v33, v32, s6
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; kill: killed $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr33
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; kill: killed $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:  .LBB96_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB96_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v18, 0x300
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v61
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v60
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v58
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v56
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v44
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v40
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v52
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; VI-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(13)
; VI-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; VI-NEXT:    s_waitcnt vmcnt(12)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    s_waitcnt vmcnt(11)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    s_waitcnt vmcnt(10)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    s_waitcnt vmcnt(9)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    s_waitcnt vmcnt(8)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    s_waitcnt vmcnt(7)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    s_waitcnt vmcnt(6)
; VI-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    s_waitcnt vmcnt(4)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v9
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v1, 0x300, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v4, v0, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v0, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v2, 0x300, v3
; VI-NEXT:    v_or_b32_e32 v0, v2, v0
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v1, v1, v4
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; VI-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_add_u16_sdwa v2, v2, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v3, 0x300, v3
; VI-NEXT:    v_or_b32_e32 v2, v3, v2
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; VI-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v3, v3, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v4, 0x300, v4
; VI-NEXT:    v_or_b32_e32 v3, v4, v3
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; VI-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v4, v4, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v5, 0x300, v5
; VI-NEXT:    v_or_b32_e32 v4, v5, v4
; VI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; VI-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v5, v5, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v6, 0x300, v6
; VI-NEXT:    v_or_b32_e32 v5, v6, v5
; VI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; VI-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v6, v6, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v7, 0x300, v7
; VI-NEXT:    v_or_b32_e32 v6, v7, v6
; VI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; VI-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v7, v7, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v8, 0x300, v8
; VI-NEXT:    v_or_b32_e32 v7, v8, v7
; VI-NEXT:    v_add_u16_e32 v8, 3, v39
; VI-NEXT:    v_lshlrev_b16_e32 v9, 8, v62
; VI-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v9, 3, v38
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v8, v8, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v9, 0x300, v9
; VI-NEXT:    v_or_b32_e32 v8, v9, v8
; VI-NEXT:    v_add_u16_e32 v9, 3, v37
; VI-NEXT:    v_lshlrev_b16_e32 v10, 8, v36
; VI-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v10, 3, v35
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v9, v9, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v10, 0x300, v10
; VI-NEXT:    v_or_b32_e32 v9, v10, v9
; VI-NEXT:    v_add_u16_e32 v10, 3, v59
; VI-NEXT:    v_lshlrev_b16_e32 v11, 8, v34
; VI-NEXT:    v_or_b32_sdwa v10, v11, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v11, 3, v33
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v10, v10, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v11, 0x300, v11
; VI-NEXT:    v_or_b32_e32 v10, v11, v10
; VI-NEXT:    v_add_u16_e32 v11, 3, v57
; VI-NEXT:    v_lshlrev_b16_e32 v12, 8, v32
; VI-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v12, 3, v63
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v11, v11, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v12, 0x300, v12
; VI-NEXT:    v_or_b32_e32 v11, v12, v11
; VI-NEXT:    v_add_u16_e32 v12, 3, v47
; VI-NEXT:    v_lshlrev_b16_e32 v13, 8, v46
; VI-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v13, 3, v45
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v12, v12, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v13, 0x300, v13
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    v_add_u16_e32 v13, 3, v43
; VI-NEXT:    v_lshlrev_b16_e32 v14, 8, v42
; VI-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v14, 3, v41
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v13, v13, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v14, 0x300, v14
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_add_u16_e32 v14, 3, v55
; VI-NEXT:    v_lshlrev_b16_e32 v15, 8, v54
; VI-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v15, 3, v53
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v14, v14, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v15, 0x300, v15
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_add_u16_e32 v15, 3, v51
; VI-NEXT:    v_lshlrev_b16_e32 v16, 8, v50
; VI-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v16, 3, v49
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v15, v15, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 0x300, v16
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v17, 8, v17
; VI-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_or_b32_sdwa v17, v19, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v17, 0x300, v17
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v21, v16, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v16, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v19, 0x300, v20
; VI-NEXT:    v_or_b32_e32 v16, v19, v16
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v17, v17, v21
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v48
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v20, 8, v20
; VI-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_sdwa v31, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v31, 0x300, v31
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v21, 8, v21
; VI-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v30, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v30, 0x300, v30
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v33, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v33, v33, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v29, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v29, 0x300, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v35, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v35, v35, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v28, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v28, 0x300, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v35
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v37, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v37, v37, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v27, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v27, 0x300, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v37
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v38, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v38, v38, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v26, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v26, 0x300, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v38
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v39, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v39, v39, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v25, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v25, 0x300, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v39
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v48, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v48, v48, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v24, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v24, 0x300, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v48
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v49, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v49, v49, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v23, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v23, 0x300, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v49
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v50, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v50, v50, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v22, 8, v22
; VI-NEXT:    v_or_b32_sdwa v22, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v22, 0x300, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v50
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_sdwa v51, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v51, v51, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v32, 8, v32
; VI-NEXT:    v_or_b32_sdwa v21, v32, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v21, 0x300, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v51
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    v_or_b32_sdwa v52, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v36, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_e32 v36, 0x300, v36
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v53, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v54, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v32, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v34, 8, v34
; VI-NEXT:    v_or_b32_sdwa v55, v34, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_sdwa v32, v19, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v34, v20, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v20, v52, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_u16_sdwa v19, v53, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u16_sdwa v18, v55, v18 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v36, v20
; VI-NEXT:    v_or_b32_e32 v30, v30, v34
; VI-NEXT:    v_or_b32_e32 v31, v31, v32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u16_e32 v52, 3, v52
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b16_e32 v53, 8, v53
; VI-NEXT:    v_or_b32_sdwa v52, v53, v52 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; VI-NEXT:    v_add_u16_e32 v52, 0x300, v52
; VI-NEXT:    v_or_b32_e32 v18, v52, v18
; VI-NEXT:    v_add_u16_e32 v52, 0x300, v54
; VI-NEXT:    v_or_b32_e32 v19, v52, v19
; VI-NEXT:  .LBB96_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v128i8_to_v64i16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:388
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:264
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:272
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:288
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:296
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:304
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:320
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
; GFX9-NEXT:    s_waitcnt vmcnt(27)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
; GFX9-NEXT:    s_waitcnt vmcnt(26)
; GFX9-NEXT:    v_lshlrev_b16_e32 v53, 8, v1
; GFX9-NEXT:    s_waitcnt vmcnt(25)
; GFX9-NEXT:    v_lshlrev_b16_e32 v51, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(24)
; GFX9-NEXT:    v_lshlrev_b16_e32 v49, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(23)
; GFX9-NEXT:    v_lshlrev_b16_e32 v39, 8, v4
; GFX9-NEXT:    s_waitcnt vmcnt(22)
; GFX9-NEXT:    v_lshlrev_b16_e32 v38, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_lshlrev_b16_e32 v34, 8, v6
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:292
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:100
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:92
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:312
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v7
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:836 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v8
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:832 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v9
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:828 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v10
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:820 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v11
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:816 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v12
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:812 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v13
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v14
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v15
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v16
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v17
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v18
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:336
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:352
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:332
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:348
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:344
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:360
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:368
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:384
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:356
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:760 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v1
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:716 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v0, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:700 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:364
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:756 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:116
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:720 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:260
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:692 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:380
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:824 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:696 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:340
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:712 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:12
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:372
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:744 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:788 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:252
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:796 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:804 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:808 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:772 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:780 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:792 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:800 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:752 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:764 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:776 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:784 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:732 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:740 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:768 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:724 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:728 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:184
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:736 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:180
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:748 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:176
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:172
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:704 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:376
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:168
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:708 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:164
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:268
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:280
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB96_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0xc0c0004
; GFX9-NEXT:    s_mov_b32 s7, 0x5040100
; GFX9-NEXT:    v_or_b32_sdwa v9, v59, v51 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v10, v47, v45, s6
; GFX9-NEXT:    v_or_b32_sdwa v11, v40, v38 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v12, v55, v54, s6
; GFX9-NEXT:    v_perm_b32 v17, v43, v42, s6
; GFX9-NEXT:    v_perm_b32 v18, v58, v57, s6
; GFX9-NEXT:    v_perm_b32 v19, v63, v46, s6
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s6
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_or_b32_sdwa v13, v48, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_perm_b32 v1, v3, v2, s7
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s7
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s7
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s7
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s7
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s7
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v33, s6
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s7
; GFX9-NEXT:    v_or_b32_sdwa v8, v60, v53 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s7
; GFX9-NEXT:    v_or_b32_sdwa v9, v44, v49 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v9, v9, v10, s7
; GFX9-NEXT:    v_or_b32_sdwa v10, v41, v39 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s7
; GFX9-NEXT:    v_or_b32_sdwa v11, v52, v34 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v11, v11, v12, s7
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v12, v50, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s7
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v13, v13, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v14, v37, v36, s6
; GFX9-NEXT:    v_perm_b32 v13, v13, v14, s7
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v14, v14, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v15, v15, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s7
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_or_b32_sdwa v15, v62, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v35, s6
; GFX9-NEXT:    v_perm_b32 v15, v15, v16, s7
; GFX9-NEXT:    v_perm_b32 v16, v32, v56, s6
; GFX9-NEXT:    v_perm_b32 v16, v17, v16, s7
; GFX9-NEXT:    v_perm_b32 v17, v19, v18, s7
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s6
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v19, v20, s6
; GFX9-NEXT:    v_perm_b32 v18, v19, v18, s7
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s6
; GFX9-NEXT:    v_perm_b32 v19, v20, v19, s7
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v20, v20, v61, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    v_perm_b32 v20, v21, v20, s7
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; kill: killed $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s6
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    v_perm_b32 v21, v22, v21, s7
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s6
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    v_perm_b32 v22, v23, v22, s7
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s6
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s6
; GFX9-NEXT:    v_perm_b32 v23, v24, v23, s7
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v24, v24, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v24, v25, v24, s7
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v25, v25, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s6
; GFX9-NEXT:    v_perm_b32 v25, v25, v26, s7
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v26, v26, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v26, v27, v26, s7
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v27, v27, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s6
; GFX9-NEXT:    v_perm_b32 v27, v27, v28, s7
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v28, v28, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v28, v29, v28, s7
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v29, v29, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s6
; GFX9-NEXT:    v_perm_b32 v29, v29, v30, s7
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v30, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_perm_b32 v30, v31, v30, s7
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v31, v31, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v32, v33, v32, s6
; GFX9-NEXT:    v_perm_b32 v31, v31, v32, s7
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    ; kill: killed $vgpr32
; GFX9-NEXT:    ; implicit-def: $vgpr32
; GFX9-NEXT:  .LBB96_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB96_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b32 s6, 0x5040100
; GFX9-NEXT:    v_lshlrev_b16_e32 v9, 8, v33
; GFX9-NEXT:    v_lshlrev_b16_e32 v10, 8, v45
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v40
; GFX9-NEXT:    v_or_b32_sdwa v11, v38, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_lshlrev_b16_e32 v12, 8, v54
; GFX9-NEXT:    v_lshlrev_b16_e32 v18, 8, v46
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v56
; GFX9-NEXT:    v_lshlrev_b16_e32 v20, 8, v42
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:836 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:832 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:828 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:816 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:812 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:760 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:716 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(21)
; GFX9-NEXT:    v_add_u16_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v1, 8, v1
; GFX9-NEXT:    v_or_b32_sdwa v0, v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(20)
; GFX9-NEXT:    v_lshlrev_b16_e32 v2, 8, v2
; GFX9-NEXT:    s_waitcnt vmcnt(19)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_lshlrev_b16_e32 v23, 8, v23
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v27, 8, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v1, 3, v1
; GFX9-NEXT:    v_or_b32_sdwa v1, v2, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v1, 0x300, v1
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v0
; GFX9-NEXT:    v_add_u16_e32 v0, 0x300, v2
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v3
; GFX9-NEXT:    v_perm_b32 v0, v2, v0, s6
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v1, v1, v4, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v2, 3, v2
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v3, 8, v3
; GFX9-NEXT:    v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_add_u16_e32 v2, 0x300, v2
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s6
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v3, 3, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v4, 8, v4
; GFX9-NEXT:    v_or_b32_sdwa v3, v4, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v3, 0x300, v3
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s6
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v4, 3, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v5, 8, v5
; GFX9-NEXT:    v_or_b32_sdwa v4, v5, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v4, 0x300, v4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    v_perm_b32 v4, v5, v4, s6
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v5, 3, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v6, 8, v6
; GFX9-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v5, 0x300, v5
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s6
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v6, 3, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v7, 8, v7
; GFX9-NEXT:    v_or_b32_sdwa v6, v7, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v6, 0x300, v6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    v_perm_b32 v6, v7, v6, s6
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v7, 3, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v8, 8, v8
; GFX9-NEXT:    v_or_b32_sdwa v7, v8, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v7, 0x300, v7
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v8
; GFX9-NEXT:    v_or_b32_sdwa v8, v9, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s6
; GFX9-NEXT:    v_add_u16_e32 v8, 3, v60
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v59
; GFX9-NEXT:    v_or_b32_sdwa v8, v53, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_or_b32_sdwa v9, v51, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v8, 0x300, v8
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_perm_b32 v8, v9, v8, s6
; GFX9-NEXT:    v_add_u16_e32 v9, 3, v47
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v44
; GFX9-NEXT:    v_or_b32_sdwa v10, v49, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v9, 0x300, v9
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s6
; GFX9-NEXT:    v_add_u16_e32 v10, 3, v41
; GFX9-NEXT:    v_or_b32_sdwa v10, v39, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v10, 0x300, v10
; GFX9-NEXT:    v_perm_b32 v10, v11, v10, s6
; GFX9-NEXT:    v_add_u16_e32 v11, 3, v55
; GFX9-NEXT:    v_or_b32_sdwa v11, v12, v11 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v52
; GFX9-NEXT:    v_or_b32_sdwa v12, v34, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v11, 0x300, v11
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s6
; GFX9-NEXT:    v_add_u16_e32 v12, 3, v50
; GFX9-NEXT:    v_or_b32_sdwa v12, v13, v12 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v48
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v12, 0x300, v12
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    v_perm_b32 v12, v13, v12, s6
; GFX9-NEXT:    v_add_u16_e32 v13, 3, v37
; GFX9-NEXT:    v_lshlrev_b16_e32 v14, 8, v36
; GFX9-NEXT:    v_or_b32_sdwa v13, v14, v13 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:720 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v13, 0x300, v13
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v14
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s6
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:820 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v14, 3, v14
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v14, v15, v14 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v14, 0x300, v14
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_perm_b32 v14, v15, v14, s6
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:824 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v16, 8, v35
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v15, 3, v15
; GFX9-NEXT:    v_or_b32_sdwa v15, v16, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v62
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v15, 0x300, v15
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v16
; GFX9-NEXT:    v_perm_b32 v15, v16, v15, s6
; GFX9-NEXT:    v_add_u16_e32 v16, 3, v58
; GFX9-NEXT:    v_lshlrev_b16_e32 v17, 8, v57
; GFX9-NEXT:    v_or_b32_sdwa v16, v17, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v17, 3, v63
; GFX9-NEXT:    v_or_b32_sdwa v17, v18, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v32
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v43
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v16
; GFX9-NEXT:    v_add_u16_e32 v16, 0x300, v18
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v19
; GFX9-NEXT:    v_perm_b32 v16, v18, v16, s6
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:744 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:708 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v17, 0x300, v17
; GFX9-NEXT:    v_perm_b32 v17, v17, v20, s6
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:756 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u16_e32 v18, 3, v18
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b16_e32 v19, 8, v19
; GFX9-NEXT:    v_or_b32_sdwa v18, v19, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:704 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v31, 0x300, v18
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v19, 3, v19
; GFX9-NEXT:    v_or_b32_sdwa v19, v20, v19 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:700 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v32, 0x300, v19
; GFX9-NEXT:    v_perm_b32 v31, v32, v31, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v20, 3, v20
; GFX9-NEXT:    v_or_b32_sdwa v20, v21, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:692 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v30, 0x300, v20
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v21, 3, v21
; GFX9-NEXT:    v_or_b32_sdwa v21, v22, v21 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:712 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v33, 0x300, v21
; GFX9-NEXT:    v_perm_b32 v30, v33, v30, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v22, 3, v22
; GFX9-NEXT:    v_or_b32_sdwa v22, v23, v22 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v29, 0x300, v22
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v23, 3, v23
; GFX9-NEXT:    v_or_b32_sdwa v23, v24, v23 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v34, 0x300, v23
; GFX9-NEXT:    v_perm_b32 v29, v34, v29, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v24, 3, v24
; GFX9-NEXT:    v_or_b32_sdwa v24, v25, v24 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v25, 3, v25
; GFX9-NEXT:    v_or_b32_sdwa v25, v26, v25 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:696 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v35, 0x300, v25
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v26, 3, v26
; GFX9-NEXT:    v_or_b32_sdwa v26, v27, v26 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v36, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v36, 0x300, v36
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v37, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v38, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v39, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v25, 0x300, v39
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v48, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v49, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v50, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:808 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:804 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v39, 0x300, v50
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v51, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:796 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:788 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v23, 0x300, v51
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v52, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:800 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:792 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v53, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:780 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:772 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v22, 0x300, v53
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v54, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:784 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:776 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v55, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:764 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:752 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v21, 0x300, v55
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v40, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:768 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v61
; GFX9-NEXT:    v_add_u16_e32 v50, 0x300, v40
; GFX9-NEXT:    v_perm_b32 v21, v50, v21, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    v_or_b32_sdwa v41, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:740 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:732 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v20, 0x300, v41
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v42, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:748 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:736 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v51, 0x300, v42
; GFX9-NEXT:    v_perm_b32 v20, v51, v20, s6
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v43, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:728 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:724 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v19, 0x300, v43
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v44, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v45, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u16_e32 v18, 0x300, v45
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u16_e32 v27, 3, v27
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b16_e32 v28, 8, v28
; GFX9-NEXT:    v_or_b32_sdwa v46, v28, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_0
; GFX9-NEXT:    v_add_u16_e32 v28, 0x300, v24
; GFX9-NEXT:    v_add_u16_e32 v27, 0x300, v26
; GFX9-NEXT:    v_add_u16_e32 v26, 0x300, v37
; GFX9-NEXT:    v_add_u16_e32 v37, 0x300, v38
; GFX9-NEXT:    v_add_u16_e32 v38, 0x300, v48
; GFX9-NEXT:    v_add_u16_e32 v24, 0x300, v49
; GFX9-NEXT:    v_add_u16_e32 v48, 0x300, v52
; GFX9-NEXT:    v_add_u16_e32 v49, 0x300, v54
; GFX9-NEXT:    v_add_u16_e32 v52, 0x300, v44
; GFX9-NEXT:    v_add_u16_e32 v53, 0x300, v46
; GFX9-NEXT:    v_perm_b32 v18, v53, v18, s6
; GFX9-NEXT:    v_perm_b32 v19, v52, v19, s6
; GFX9-NEXT:    v_perm_b32 v22, v49, v22, s6
; GFX9-NEXT:    v_perm_b32 v23, v48, v23, s6
; GFX9-NEXT:    v_perm_b32 v24, v39, v24, s6
; GFX9-NEXT:    v_perm_b32 v25, v38, v25, s6
; GFX9-NEXT:    v_perm_b32 v26, v37, v26, s6
; GFX9-NEXT:    v_perm_b32 v27, v36, v27, s6
; GFX9-NEXT:    v_perm_b32 v28, v35, v28, s6
; GFX9-NEXT:  .LBB96_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64i16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:488
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v72, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v73, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v74, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v75, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v76, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v77, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v78, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v79, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v88, s32 offset:392
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v31, off, s32 offset:384
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:380
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v75, off, s32 offset:376
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v78, off, s32 offset:372
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v31, off, s32 offset:368
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:364
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v70, off, s32 offset:360
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v39, off, s32 offset:356
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:352
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:348
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v46, off, s32 offset:344
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v74, off, s32 offset:340
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:336
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v35, off, s32 offset:332
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v69, off, s32 offset:328
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v38, off, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v68, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v61, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v72, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v36, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v67, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v183, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v59, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v66, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v32, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v33, off, s32 offset:260
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v44, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v43, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:188
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:132
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v65, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:388
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v52, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v53, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v54, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v55, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v64, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v71, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v51, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v76, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v79, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v50, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v48, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v73, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v77, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v49, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v37, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v56, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v63, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_hi_b16 v34, off, s32 offset:12
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v41, off, s32
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v60.l, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v29.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v58.l, v28.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v182.l, v27.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v57.l, v26.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v25.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v62.l, v24.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v178.l, v23.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v22.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v166.l, v21.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v20.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v162.l, v19.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v181.l, v18.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v133.l, v17.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v164.l, v16.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v135.l, v15.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v165.l, v14.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v132.l, v13.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v177.l, v12.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v114.l, v11.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v146.l, v10.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v128.l, v9.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v147.l, v8.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v97.l, v7.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v130.l, v6.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v101.l, v5.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v160.l, v4.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v100.l, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v117.l, v2.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v87.l, v1.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v113.l, v0.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(62)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v69.l, 8, v69.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v70.l, 8, v70.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v69.h, 8, v69.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v68.h, 8, v68.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.l, 8, v67.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v68.l, 8, v68.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v67.h, 8, v67.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.h, 8, v66.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v66.l, 8, v66.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(33)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v65.h, 8, v65.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(29)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v88
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(27)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.l, 8, v52.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v52.h, 8, v52.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(25)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v53.l, 8, v53.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(24)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.l, 8, v53.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v53.h, 8, v54.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(22)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.l, 8, v55.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v55.h, 8, v55.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(20)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v54.h, 8, v64.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v65.l, 8, v64.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(18)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v64.l, 8, v71.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v64.h, 8, v71.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.l, 8, v70.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v70.h, 8, v31.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v71.h, 8, v31.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB96_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v130, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v165, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v45, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v60, v41, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v9.l, 0xff, v37.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v160, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v11.h, 0xff, v48.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v3.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v177, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v13.h, 0xff, v51.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v40, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v14.h, 0xff, v32.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v7.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v58, v42, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v9.l, v53.l
; GFX11-TRUE16-NEXT:    v_and_b16 v10.h, 0xff, v49.h
; GFX11-TRUE16-NEXT:    v_and_b16 v11.l, 0xff, v38.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v11.h, v55.l
; GFX11-TRUE16-NEXT:    v_and_b16 v12.h, 0xff, v50.h
; GFX11-TRUE16-NEXT:    v_and_b16 v13.l, 0xff, v49.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v13.h, v65.l
; GFX11-TRUE16-NEXT:    v_and_b16 v15.l, 0xff, v51.h
; GFX11-TRUE16-NEXT:    v_and_b16 v15.h, 0xff, v50.l
; GFX11-TRUE16-NEXT:    v_or_b16 v16.l, v14.h, v65.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v63, v56, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v98, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v10.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.l, v10.h, v54.l
; GFX11-TRUE16-NEXT:    v_or_b16 v10.h, v11.l, v53.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v77, v73, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v102, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v12.l
; GFX11-TRUE16-NEXT:    v_or_b16 v12.l, v12.h, v55.h
; GFX11-TRUE16-NEXT:    v_or_b16 v12.h, v13.l, v54.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v79, v76, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v119, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v14.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.l, v15.l, v64.l
; GFX11-TRUE16-NEXT:    v_or_b16 v14.h, v15.h, v64.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v83, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v144, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v16.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v96, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v149, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v17.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v99, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v163, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v18.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v103, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v179, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v117, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v19.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v148, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v44, v176, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v25.l, 0xff, v34.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v146, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v20.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v145, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v27.h, 0xff, v35.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v181, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v21.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v161, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v29.h, 0xff, v39.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v57, v182, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v22.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v43, v167, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_and_b16 v30.l, 0xff, v48.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v113, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v23.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v180, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v2.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v147, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v26.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v25.l, v67.h
; GFX11-TRUE16-NEXT:    v_and_b16 v26.h, 0xff, v36.h
; GFX11-TRUE16-NEXT:    v_and_b16 v27.l, 0xff, v33.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v4.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v164, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v27.h, v68.h
; GFX11-TRUE16-NEXT:    v_and_b16 v28.h, 0xff, v38.h
; GFX11-TRUE16-NEXT:    v_and_b16 v29.l, 0xff, v35.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v62, v47, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v29.h, v69.l
; GFX11-TRUE16-NEXT:    v_and_b16 v31.l, 0xff, v39.h
; GFX11-TRUE16-NEXT:    v_and_b16 v31.h, 0xff, v36.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.l, 0xff, v37.l
; GFX11-TRUE16-NEXT:    v_and_b16 v8.h, 0xff, v34.h
; GFX11-TRUE16-NEXT:    v_and_b16 v24.l, 0xff, v33.h
; GFX11-TRUE16-NEXT:    v_and_b16 v24.h, 0xff, v32.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v30.l, v71.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v59, v183, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v26.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.l, v26.h, v68.l
; GFX11-TRUE16-NEXT:    v_or_b16 v26.h, v27.l, v67.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v72, v61, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v28.l
; GFX11-TRUE16-NEXT:    v_or_b16 v28.l, v28.h, v69.h
; GFX11-TRUE16-NEXT:    v_or_b16 v28.h, v29.l, v70.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v74, v46, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v30.h
; GFX11-TRUE16-NEXT:    v_or_b16 v30.l, v31.l, v71.l
; GFX11-TRUE16-NEXT:    v_or_b16 v30.h, v31.h, v70.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v78, v75, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b16 v8.l, v8.l, v52.l
; GFX11-TRUE16-NEXT:    v_or_b16 v8.h, v8.h, v52.h
; GFX11-TRUE16-NEXT:    v_or_b16 v24.l, v24.l, v66.l
; GFX11-TRUE16-NEXT:    v_or_b16 v24.h, v24.h, v66.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v32.l
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr177_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr181_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr40_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr45_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr178_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr62_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr47_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr57_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr182_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr58_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr42_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr60_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr41_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr63_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr56_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr77_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr73_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr79_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr76_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr50_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr43_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr167_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr179_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr180_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr44_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr176_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr59_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr183_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr72_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr61_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr74_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr46_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr39_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr78_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr75_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr53_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr55_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr67_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_hi16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr71_hi16
; GFX11-TRUE16-NEXT:  .LBB96_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB96_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v78.l, 3
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(10)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v48.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, v39.h, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v75.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v36.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v74.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v43.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v71.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v71.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v39.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v31.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v70.h, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.l, 8, v46.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v38.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v35.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v30.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.l, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v69.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v72.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v29.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v69.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v70.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v61.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v35.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v36.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v28.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v33.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v59.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v68.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v68.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v34.l, 3
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v183.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v27.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v67.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.h, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v33.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v32.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v26.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v67.h, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v180.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v25.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v44.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v66.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v151.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v66.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v176.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v24.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v167.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v179.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v161.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v23.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v150.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v134.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v163.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v145.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v131.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v118.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v149.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v22.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v129.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v148.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v21.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v144.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v115.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v116.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v103.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v20.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v119.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v86.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v112.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v99.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v19.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.h, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v85.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v102.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v96.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v18.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v84.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v82.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v98.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v83.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v81.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v32.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v80.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v17.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v3.l, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v51.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v50.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v16.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v65.h, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v79.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v15.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v51.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v64.l, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v64.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v76.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v50.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v14.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v65.l, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v49.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v77.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v13.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v55.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v48.h, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v73.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v49.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v54.h, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v2.l, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v38.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v12.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v55.l, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v54.l, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v63.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v11.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v37.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v53.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v56.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v37.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v34.h, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v10.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v53.l, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v58.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v9.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v52.l, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v52.h, v1.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v42.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v60.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v62.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v8.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v41.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v47.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v57.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v40.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v182.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v166.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, v45.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v7.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.h, v0.h
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v3.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v178.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v164.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v6.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, v181.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.h, 8, v133.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v0.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.l, 8, v162.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v5.h, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v177.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v1.h, v1.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v165.l, 3
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v2.l, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, v147.l, 3
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v2.h, 8, v132.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.l, 8, v135.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.l, 0xff, v2.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v3.h, 8, v128.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.l, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v2.h, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.l, v3.l, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v4.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v3.h, v2.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v146.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, v160.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, v113.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.l, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.l, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, v130.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v32.h, v117.l, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v3.h, 0x300, v1.l
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v1.l, 8, v114.l
; GFX11-TRUE16-NEXT:    v_and_b16 v0.h, 0xff, v0.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v32.l, 8, v101.l
; GFX11-TRUE16-NEXT:    v_and_b16 v1.h, 0xff, v1.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.l, 8, v97.l
; GFX11-TRUE16-NEXT:    v_and_b16 v2.h, 0xff, v2.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v33.h, 8, v87.l
; GFX11-TRUE16-NEXT:    v_and_b16 v32.h, 0xff, v32.h
; GFX11-TRUE16-NEXT:    v_lshlrev_b16 v34.l, 8, v100.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.l, v1.l, v0.l
; GFX11-TRUE16-NEXT:    v_or_b16 v0.h, v32.l, v0.h
; GFX11-TRUE16-NEXT:    v_or_b16 v1.h, v33.l, v1.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.l, v33.h, v2.h
; GFX11-TRUE16-NEXT:    v_or_b16 v32.h, v34.l, v32.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v2.h, 0x300, v0.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.l, 0x300, v0.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v1.h, 0x300, v1.h
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.l, 0x300, v32.l
; GFX11-TRUE16-NEXT:    v_add_nc_u16 v0.h, 0x300, v32.h
; GFX11-TRUE16-NEXT:  .LBB96_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_clause 0x18 ; 100-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v88, off, s32 offset:392
; GFX11-TRUE16-NEXT:    scratch_load_b32 v79, off, s32 offset:396
; GFX11-TRUE16-NEXT:    scratch_load_b32 v78, off, s32 offset:400
; GFX11-TRUE16-NEXT:    scratch_load_b32 v77, off, s32 offset:404
; GFX11-TRUE16-NEXT:    scratch_load_b32 v76, off, s32 offset:408
; GFX11-TRUE16-NEXT:    scratch_load_b32 v75, off, s32 offset:412
; GFX11-TRUE16-NEXT:    scratch_load_b32 v74, off, s32 offset:416
; GFX11-TRUE16-NEXT:    scratch_load_b32 v73, off, s32 offset:420
; GFX11-TRUE16-NEXT:    scratch_load_b32 v72, off, s32 offset:424
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:428
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:432
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:436
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:440
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:444
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:448
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:452
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:456
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:460
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:464
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:468
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:472
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:476
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:480
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:484
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:488
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64i16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:580
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:516
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v77, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v78, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v79, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v88, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v89, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v90, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v91, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v92, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v93, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v94, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v95, s32 offset:456
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v104, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v105, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v106, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v107, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v108, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v109, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v110, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v111, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v120, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v121, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v122, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v123, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v124, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v125, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v126, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v127, s32 offset:392
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v84, v24 :: v_dual_mov_b32 v51, v19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v66, v23 :: v_dual_mov_b32 v71, v20
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v70, v22 :: v_dual_mov_b32 v67, v17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v68, v21 :: v_dual_mov_b32 v65, v18
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v69, v16 :: v_dual_mov_b32 v50, v15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v52, v14 :: v_dual_mov_b32 v53, v9
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v54, v13 :: v_dual_mov_b32 v49, v10
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v64, v12 :: v_dual_mov_b32 v33, v7
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v32, v11 :: v_dual_mov_b32 v55, v8
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v39, v6 :: v_dual_mov_b32 v36, v5
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v48, v4 :: v_dual_mov_b32 v37, v2
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v34, v3 :: v_dual_mov_b32 v35, v1
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v38, v0
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v0, off, s32 offset:384
; GFX11-FAKE16-NEXT:    scratch_load_u16 v105, off, s32 offset:380
; GFX11-FAKE16-NEXT:    scratch_load_u16 v93, off, s32 offset:376
; GFX11-FAKE16-NEXT:    scratch_load_u16 v104, off, s32 offset:372
; GFX11-FAKE16-NEXT:    scratch_load_u16 v1, off, s32 offset:368
; GFX11-FAKE16-NEXT:    scratch_load_u16 v94, off, s32 offset:364
; GFX11-FAKE16-NEXT:    scratch_load_u16 v2, off, s32 offset:360
; GFX11-FAKE16-NEXT:    scratch_load_u16 v106, off, s32 offset:356
; GFX11-FAKE16-NEXT:    scratch_load_u16 v3, off, s32 offset:352
; GFX11-FAKE16-NEXT:    scratch_load_u16 v91, off, s32 offset:348
; GFX11-FAKE16-NEXT:    scratch_load_u16 v77, off, s32 offset:344
; GFX11-FAKE16-NEXT:    scratch_load_u16 v89, off, s32 offset:340
; GFX11-FAKE16-NEXT:    scratch_load_u16 v4, off, s32 offset:336
; GFX11-FAKE16-NEXT:    scratch_load_u16 v88, off, s32 offset:332
; GFX11-FAKE16-NEXT:    scratch_load_u16 v5, off, s32 offset:328
; GFX11-FAKE16-NEXT:    scratch_load_u16 v92, off, s32 offset:324
; GFX11-FAKE16-NEXT:    scratch_load_u16 v6, off, s32 offset:320
; GFX11-FAKE16-NEXT:    scratch_load_u16 v78, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v62, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v76, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v7, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v73, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v8, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v79, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v9, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v63, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v45, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v61, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v10, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v58, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v11, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v72, off, s32 offset:260
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v41, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v44, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v57, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v60, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v40, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v42, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v43, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v183, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:188
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:132
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v12, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_b32 v13, off, s32 offset:388
; GFX11-FAKE16-NEXT:    scratch_load_u16 v14, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v15, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v16, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v17, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v18, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v19, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v20, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v21, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v22, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v23, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v24, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:12
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v83, v30 :: v_dual_mov_b32 v86, v28
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v85, v29 :: v_dual_mov_b32 v80, v27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v82, v26 :: v_dual_mov_b32 v81, v25
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(62)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v127, 8, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v126, 8, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v125, 8, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v123, 8, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v122, 8, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v124, 8, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v121, 8, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v120, 8, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v111, 8, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v110, 8, v9
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v108, 8, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v109, 8, v11
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(33)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v107, 8, v12
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(29)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v13
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(28)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v182, 8, v14
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(27)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v166, 8, v15
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(26)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v180, 8, v16
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(25)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v46, 8, v17
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(24)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v47, 8, v18
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(23)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v56, 8, v19
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(22)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v75, 8, v20
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(21)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v59, 8, v21
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(20)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v74, 8, v22
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(19)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v95, 8, v23
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(18)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v90, 8, v24
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB96_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v38, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v37, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v48, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v39, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v64, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v52, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v1, v0, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v69, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v3, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v55, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v49, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v65, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v84, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v82, v80, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v71, v68, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v70, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v3, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v5, v4, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v7, v6, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v11, v10, 0x5040100
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v118
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v113
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v9, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v86, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v83, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v99
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v7, v182
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v7, 0xff, v134
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v129
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v11, v180
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v112, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v166
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v7, v46
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, v13, v47
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v9, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xff, v132
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v11, v14, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v151
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v10, v12, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v13, v15, 0x5040100
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, v16, v56
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v145
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xff, v150
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v16, v11, v75
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v167
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xff, v162
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v130, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, v14, v59
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v15, v74
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v148, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v19, v11, v95
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, v17, v90
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xff, v116
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v12, v13, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v14, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v15, v18, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v17, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v20, v107
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v115, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v103, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v98, v96, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v131, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v117, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v147, v146, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v133, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v164, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v149, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v18, v17, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v20, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v21, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v24, v23, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v177, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v165, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v183, v178, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v179, v176, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v43, v42, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v40, v181, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v60, v57, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v44, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v72
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v21, v20, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v23, v22, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v25, v24, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v58
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v26, v28, v109
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v79
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v73
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xff, v63
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v24, v24, v108
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v61, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v27, v111
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v28, v120
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v25, v110
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v24, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xff, v91
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v28, v27, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v88
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v25, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xff, v92
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xff, v94
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xff, v105
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v27, v122
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xff, v106
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v28, v30, v121
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v76, v62, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v29, v124
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v31, v123
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v89, v77, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, v27, v125
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v126
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v34, v127
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v104, v93, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v28, v30, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v31, v35, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v33, v36, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v34, v37, 0x5040100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr39
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr55
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr32
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr51
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr68
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr79
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr78
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr92
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr88
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr89
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr77
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr91
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr106
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr94
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr104
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr93
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr105
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr95
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr90
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr107
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr109
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr108
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr110
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr111
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr120
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr121
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr124
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr122
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr123
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr125
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr126
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr127
; GFX11-FAKE16-NEXT:  .LBB96_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB96_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v104, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v105, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v2, 8, v93
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v106, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v94, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v92, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v2, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v127, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v89, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v125, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, v126, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v31, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v89, 0x300, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v1, 8, v77
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v91, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v30, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v77, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v88, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v1, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v76, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v62
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v29, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v123, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v124, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v78, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v62, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v28, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v122, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v5, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v79, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v73, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v73, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v27, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v121, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v61, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v72, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v61, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v63, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v120, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v111, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v45
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v45, 0x300, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v26, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v110, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v58, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v60, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v42
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v25, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v58, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v109, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v57
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v44, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v24, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v43, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v108, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v41
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v40, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v40, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v23, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v183, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v181, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v179, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v22, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v178
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v176
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v177, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v176, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v165, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v163
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v161
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v164, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v21, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v161, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v160
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v149, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v20, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v147, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v147, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v144
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v133, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v19, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v131, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v119
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v128
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, v117, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v117, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v18, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v5, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v6
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v103, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v102, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v98, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v17, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v97
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v115, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v96, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v116, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v100
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(16)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v167, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v16, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v97, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v107, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v162, 3
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(13)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v148, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v15, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v98, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v95, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v150, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v135
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(11)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v151, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v14, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v90, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v4, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v145, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v100, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v74, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v13, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v75, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(8)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v130, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v103, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v132, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v12, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v59, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v114
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(6)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v134, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v129, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v114, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v56, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v5
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v112, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v11, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v112, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v46, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v47, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v113, 3
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v118, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v10, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v101, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v99, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v86, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v9, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v180, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v182, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, v83, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v83, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v8, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v166, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v4, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v5
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v87
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v84, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v84, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v7, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v0, v82, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v81
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v4, 8, v80
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v80, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v71, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v3, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v70, 3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v5, 8, v68
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v4, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v6, 8, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, v69, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v68, 0x300, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v5, v1
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v66, 8, v67
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v6, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v65, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v5, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v1, v64, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v6, 0x300, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v3, 8, v54
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v52, v52, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v1, 0xff, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v67, 0x300, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v66, v4
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v51, 8, v51
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v50, 8, v50
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v3, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v52
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v4, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, v55, 3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v0, v51, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v49, v49, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v51, 0x300, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, v50, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, v48, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v39, v39, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v38, v38, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v37, v37, 3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xff, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v53, 8, v53
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xff, v49
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xff, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v36, 8, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xff, v39
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xff, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v35, 8, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xff, v37
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v34, 8, v34
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v33, 8, v33
; GFX11-FAKE16-NEXT:    v_lshlrev_b16 v32, 8, v32
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, v53, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, v36, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, v35, v38
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, v34, v37
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v33, v33, v39
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v32, v32, v48
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v3, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v35, 0x300, v35
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v34, 0x300, v34
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v33, 0x300, v33
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v32, 0x300, v32
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v36, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u16 v37, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v34, v35, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v33, v3, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v32, v2, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v36, v51, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v37, v4, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v6, v5, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v68, v67, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v80, v7, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v84, v8, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v83, v9, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v101, v10, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v112, v11, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v114, v12, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v103, v13, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v100, v14, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v98, v15, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v97, v16, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v96, v17, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v102, v18, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v117, v19, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v147, v20, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v161, v21, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v176, v22, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v181, v23, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v40, v24, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v58, v25, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v45, v26, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v61, v27, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v73, v28, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v62, v29, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v77, v30, 0x5040100
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v89, v31, 0x5040100
; GFX11-FAKE16-NEXT:  .LBB96_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_clause 0x1f ; 128-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v127, off, s32 offset:392
; GFX11-FAKE16-NEXT:    scratch_load_b32 v126, off, s32 offset:396
; GFX11-FAKE16-NEXT:    scratch_load_b32 v125, off, s32 offset:400
; GFX11-FAKE16-NEXT:    scratch_load_b32 v124, off, s32 offset:404
; GFX11-FAKE16-NEXT:    scratch_load_b32 v123, off, s32 offset:408
; GFX11-FAKE16-NEXT:    scratch_load_b32 v122, off, s32 offset:412
; GFX11-FAKE16-NEXT:    scratch_load_b32 v121, off, s32 offset:416
; GFX11-FAKE16-NEXT:    scratch_load_b32 v120, off, s32 offset:420
; GFX11-FAKE16-NEXT:    scratch_load_b32 v111, off, s32 offset:424
; GFX11-FAKE16-NEXT:    scratch_load_b32 v110, off, s32 offset:428
; GFX11-FAKE16-NEXT:    scratch_load_b32 v109, off, s32 offset:432
; GFX11-FAKE16-NEXT:    scratch_load_b32 v108, off, s32 offset:436
; GFX11-FAKE16-NEXT:    scratch_load_b32 v107, off, s32 offset:440
; GFX11-FAKE16-NEXT:    scratch_load_b32 v106, off, s32 offset:444
; GFX11-FAKE16-NEXT:    scratch_load_b32 v105, off, s32 offset:448
; GFX11-FAKE16-NEXT:    scratch_load_b32 v104, off, s32 offset:452
; GFX11-FAKE16-NEXT:    scratch_load_b32 v95, off, s32 offset:456
; GFX11-FAKE16-NEXT:    scratch_load_b32 v94, off, s32 offset:460
; GFX11-FAKE16-NEXT:    scratch_load_b32 v93, off, s32 offset:464
; GFX11-FAKE16-NEXT:    scratch_load_b32 v92, off, s32 offset:468
; GFX11-FAKE16-NEXT:    scratch_load_b32 v91, off, s32 offset:472
; GFX11-FAKE16-NEXT:    scratch_load_b32 v90, off, s32 offset:476
; GFX11-FAKE16-NEXT:    scratch_load_b32 v89, off, s32 offset:480
; GFX11-FAKE16-NEXT:    scratch_load_b32 v88, off, s32 offset:484
; GFX11-FAKE16-NEXT:    scratch_load_b32 v79, off, s32 offset:488
; GFX11-FAKE16-NEXT:    scratch_load_b32 v78, off, s32 offset:492
; GFX11-FAKE16-NEXT:    scratch_load_b32 v77, off, s32 offset:496
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:500
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:504
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:508
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:512
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:516
; GFX11-FAKE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:520
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:524
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:528
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:532
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:536
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:540
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:544
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:548
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:552
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:556
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:560
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:564
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:568
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:572
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:576
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:580
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define inreg <64 x i16> @bitcast_v128i8_to_v64i16_scalar(<128 x i8> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v128i8_to_v64i16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    ; implicit-def: $vgpr43 : SGPR spill to VGPR lane
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_writelane_b32 v40, s30, 0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v43, s29, 0
; SI-NEXT:    v_writelane_b32 v43, s28, 1
; SI-NEXT:    v_writelane_b32 v43, s27, 2
; SI-NEXT:    v_writelane_b32 v43, s26, 3
; SI-NEXT:    v_writelane_b32 v43, s25, 4
; SI-NEXT:    v_writelane_b32 v40, s31, 1
; SI-NEXT:    v_writelane_b32 v43, s24, 5
; SI-NEXT:    v_writelane_b32 v40, s34, 2
; SI-NEXT:    v_writelane_b32 v43, s23, 6
; SI-NEXT:    v_writelane_b32 v40, s35, 3
; SI-NEXT:    v_writelane_b32 v43, s22, 7
; SI-NEXT:    v_writelane_b32 v40, s36, 4
; SI-NEXT:    v_writelane_b32 v43, s21, 8
; SI-NEXT:    v_writelane_b32 v40, s37, 5
; SI-NEXT:    v_writelane_b32 v43, s20, 9
; SI-NEXT:    v_writelane_b32 v40, s38, 6
; SI-NEXT:    v_writelane_b32 v43, s19, 10
; SI-NEXT:    v_writelane_b32 v40, s39, 7
; SI-NEXT:    v_writelane_b32 v43, s18, 11
; SI-NEXT:    v_writelane_b32 v40, s48, 8
; SI-NEXT:    v_writelane_b32 v43, s17, 12
; SI-NEXT:    v_writelane_b32 v40, s49, 9
; SI-NEXT:    v_writelane_b32 v43, s16, 13
; SI-NEXT:    v_readfirstlane_b32 s4, v30
; SI-NEXT:    v_writelane_b32 v40, s50, 10
; SI-NEXT:    v_writelane_b32 v43, s4, 14
; SI-NEXT:    v_readfirstlane_b32 s4, v29
; SI-NEXT:    v_writelane_b32 v40, s51, 11
; SI-NEXT:    v_writelane_b32 v43, s4, 15
; SI-NEXT:    v_readfirstlane_b32 s4, v28
; SI-NEXT:    v_writelane_b32 v40, s52, 12
; SI-NEXT:    v_writelane_b32 v43, s4, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v27
; SI-NEXT:    v_writelane_b32 v40, s53, 13
; SI-NEXT:    v_writelane_b32 v43, s4, 17
; SI-NEXT:    v_readfirstlane_b32 s4, v26
; SI-NEXT:    v_writelane_b32 v40, s54, 14
; SI-NEXT:    v_writelane_b32 v43, s4, 18
; SI-NEXT:    v_readfirstlane_b32 s4, v25
; SI-NEXT:    v_writelane_b32 v40, s55, 15
; SI-NEXT:    v_writelane_b32 v43, s4, 19
; SI-NEXT:    v_readfirstlane_b32 s4, v24
; SI-NEXT:    v_writelane_b32 v40, s64, 16
; SI-NEXT:    v_writelane_b32 v43, s4, 20
; SI-NEXT:    v_readfirstlane_b32 s4, v23
; SI-NEXT:    v_writelane_b32 v40, s65, 17
; SI-NEXT:    v_writelane_b32 v43, s4, 21
; SI-NEXT:    v_readfirstlane_b32 s4, v22
; SI-NEXT:    v_writelane_b32 v40, s66, 18
; SI-NEXT:    v_writelane_b32 v43, s4, 22
; SI-NEXT:    v_readfirstlane_b32 s4, v21
; SI-NEXT:    v_writelane_b32 v40, s67, 19
; SI-NEXT:    v_writelane_b32 v43, s4, 23
; SI-NEXT:    v_readfirstlane_b32 s4, v20
; SI-NEXT:    v_writelane_b32 v40, s68, 20
; SI-NEXT:    v_writelane_b32 v43, s4, 24
; SI-NEXT:    v_readfirstlane_b32 s4, v19
; SI-NEXT:    v_writelane_b32 v40, s69, 21
; SI-NEXT:    v_writelane_b32 v43, s4, 25
; SI-NEXT:    v_readfirstlane_b32 s4, v17
; SI-NEXT:    v_writelane_b32 v40, s70, 22
; SI-NEXT:    v_writelane_b32 v43, s4, 26
; SI-NEXT:    v_readfirstlane_b32 s4, v16
; SI-NEXT:    v_writelane_b32 v40, s71, 23
; SI-NEXT:    v_readfirstlane_b32 s71, v18
; SI-NEXT:    v_writelane_b32 v43, s4, 27
; SI-NEXT:    v_readfirstlane_b32 s4, v15
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:232
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:228
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:220
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:328
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:324
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:320
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:316
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:312
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:308
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:304
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:300
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:296
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:292
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:288
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:284
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:280
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:276
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:272
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:268
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:264
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:260
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:256
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:252
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:248
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:244
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:332
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:240
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:236
; SI-NEXT:    v_writelane_b32 v43, s4, 28
; SI-NEXT:    v_readfirstlane_b32 s4, v14
; SI-NEXT:    v_writelane_b32 v43, s4, 29
; SI-NEXT:    v_readfirstlane_b32 s70, v13
; SI-NEXT:    v_readfirstlane_b32 s4, v12
; SI-NEXT:    v_readfirstlane_b32 s25, v11
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:216
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:212
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:208
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:204
; SI-NEXT:    v_writelane_b32 v43, s4, 30
; SI-NEXT:    v_readfirstlane_b32 s4, v9
; SI-NEXT:    v_writelane_b32 v43, s4, 31
; SI-NEXT:    v_readfirstlane_b32 s4, v8
; SI-NEXT:    v_writelane_b32 v43, s4, 32
; SI-NEXT:    v_readfirstlane_b32 s4, v7
; SI-NEXT:    v_writelane_b32 v43, s4, 33
; SI-NEXT:    v_readfirstlane_b32 s4, v6
; SI-NEXT:    v_writelane_b32 v43, s4, 34
; SI-NEXT:    v_readfirstlane_b32 s4, v5
; SI-NEXT:    v_writelane_b32 v43, s4, 35
; SI-NEXT:    v_readfirstlane_b32 s4, v4
; SI-NEXT:    v_writelane_b32 v43, s4, 36
; SI-NEXT:    v_readfirstlane_b32 s4, v1
; SI-NEXT:    v_readfirstlane_b32 s37, v2
; SI-NEXT:    v_writelane_b32 v43, s4, 37
; SI-NEXT:    v_readfirstlane_b32 s4, v0
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176
; SI-NEXT:    v_readfirstlane_b32 s24, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:200
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:196
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:192
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:188
; SI-NEXT:    v_writelane_b32 v40, s80, 24
; SI-NEXT:    v_writelane_b32 v40, s81, 25
; SI-NEXT:    v_writelane_b32 v40, s82, 26
; SI-NEXT:    v_writelane_b32 v40, s83, 27
; SI-NEXT:    v_writelane_b32 v40, s84, 28
; SI-NEXT:    v_writelane_b32 v43, s4, 38
; SI-NEXT:    v_writelane_b32 v40, s85, 29
; SI-NEXT:    v_writelane_b32 v40, s86, 30
; SI-NEXT:    v_writelane_b32 v40, s87, 31
; SI-NEXT:    v_writelane_b32 v40, s96, 32
; SI-NEXT:    v_writelane_b32 v40, s97, 33
; SI-NEXT:    v_writelane_b32 v40, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s28, v3
; SI-NEXT:    ; implicit-def: $vgpr42 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v40, s99, 35
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s56, v19
; SI-NEXT:    v_readfirstlane_b32 s47, v20
; SI-NEXT:    v_readfirstlane_b32 s57, v21
; SI-NEXT:    v_readfirstlane_b32 s58, v22
; SI-NEXT:    v_readfirstlane_b32 s59, v23
; SI-NEXT:    v_readfirstlane_b32 s62, v24
; SI-NEXT:    v_readfirstlane_b32 s89, v25
; SI-NEXT:    v_readfirstlane_b32 s90, v26
; SI-NEXT:    v_readfirstlane_b32 s98, v27
; SI-NEXT:    v_readfirstlane_b32 s65, v28
; SI-NEXT:    v_readfirstlane_b32 s60, v29
; SI-NEXT:    v_readfirstlane_b32 s44, v30
; SI-NEXT:    v_readfirstlane_b32 s46, v31
; SI-NEXT:    v_readfirstlane_b32 s45, v32
; SI-NEXT:    v_readfirstlane_b32 s75, v33
; SI-NEXT:    v_readfirstlane_b32 s79, v34
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:172
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:168
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:160
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:156
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:152
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:148
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:144
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:136
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:132
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:128
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:124
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:120
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:116
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112
; SI-NEXT:    v_readfirstlane_b32 s96, v35
; SI-NEXT:    v_readfirstlane_b32 s84, v36
; SI-NEXT:    v_readfirstlane_b32 s80, v37
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_readfirstlane_b32 s83, v17
; SI-NEXT:    v_readfirstlane_b32 s21, v16
; SI-NEXT:    v_readfirstlane_b32 s22, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:96
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:92
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:88
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:84
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:80
; SI-NEXT:    v_readfirstlane_b32 s92, v38
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s23, v51
; SI-NEXT:    v_readfirstlane_b32 s31, v39
; SI-NEXT:    v_readfirstlane_b32 s97, v14
; SI-NEXT:    v_readfirstlane_b32 s55, v13
; SI-NEXT:    v_readfirstlane_b32 s72, v12
; SI-NEXT:    v_readfirstlane_b32 s73, v11
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:76
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:72
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:68
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:64
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:60
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:56
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:52
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:48
; SI-NEXT:    v_readfirstlane_b32 s26, v48
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v49
; SI-NEXT:    v_readfirstlane_b32 s14, v50
; SI-NEXT:    v_writelane_b32 v43, s4, 39
; SI-NEXT:    s_and_b64 s[4:5], vcc, exec
; SI-NEXT:    v_readfirstlane_b32 s82, v1
; SI-NEXT:    v_readfirstlane_b32 s94, v0
; SI-NEXT:    v_readfirstlane_b32 s63, v2
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v51, off, s[0:3], s32
; SI-NEXT:    v_readfirstlane_b32 s86, v10
; SI-NEXT:    v_readfirstlane_b32 s87, v9
; SI-NEXT:    v_readfirstlane_b32 s85, v8
; SI-NEXT:    v_readfirstlane_b32 s69, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:44
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:40
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:36
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:32
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:28
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:24
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:20
; SI-NEXT:    buffer_load_dword v50, off, s[0:3], s32 offset:16
; SI-NEXT:    v_readfirstlane_b32 s74, v3
; SI-NEXT:    v_readfirstlane_b32 s61, v4
; SI-NEXT:    v_readfirstlane_b32 s67, v5
; SI-NEXT:    v_readfirstlane_b32 s81, v6
; SI-NEXT:    v_readfirstlane_b32 s29, v19
; SI-NEXT:    v_readfirstlane_b32 s35, v20
; SI-NEXT:    v_readfirstlane_b32 s27, v21
; SI-NEXT:    v_readfirstlane_b32 s17, v22
; SI-NEXT:    v_readfirstlane_b32 s13, v23
; SI-NEXT:    v_readfirstlane_b32 s52, v24
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_readfirstlane_b32 s51, v25
; SI-NEXT:    v_readfirstlane_b32 s53, v26
; SI-NEXT:    v_readfirstlane_b32 s54, v27
; SI-NEXT:    v_readfirstlane_b32 s66, v28
; SI-NEXT:    v_readfirstlane_b32 s64, v29
; SI-NEXT:    v_readfirstlane_b32 s18, v30
; SI-NEXT:    v_readfirstlane_b32 s12, v15
; SI-NEXT:    v_readfirstlane_b32 s38, v16
; SI-NEXT:    v_readfirstlane_b32 s36, v17
; SI-NEXT:    v_readfirstlane_b32 s39, v18
; SI-NEXT:    v_readfirstlane_b32 s48, v31
; SI-NEXT:    v_readfirstlane_b32 s50, v32
; SI-NEXT:    v_writelane_b32 v42, s50, 0
; SI-NEXT:    v_readfirstlane_b32 s10, v34
; SI-NEXT:    v_writelane_b32 v42, s10, 1
; SI-NEXT:    v_writelane_b32 v42, s51, 2
; SI-NEXT:    v_readfirstlane_b32 s11, v11
; SI-NEXT:    v_writelane_b32 v42, s52, 3
; SI-NEXT:    v_writelane_b32 v42, s11, 4
; SI-NEXT:    v_writelane_b32 v42, s53, 5
; SI-NEXT:    v_writelane_b32 v42, s54, 6
; SI-NEXT:    v_writelane_b32 v42, s64, 7
; SI-NEXT:    v_writelane_b32 v42, s66, 8
; SI-NEXT:    v_writelane_b32 v42, s67, 9
; SI-NEXT:    v_writelane_b32 v42, s61, 10
; SI-NEXT:    v_writelane_b32 v42, s83, 11
; SI-NEXT:    v_writelane_b32 v42, s86, 12
; SI-NEXT:    v_writelane_b32 v42, s12, 13
; SI-NEXT:    v_writelane_b32 v42, s87, 14
; SI-NEXT:    v_writelane_b32 v42, s81, 15
; SI-NEXT:    v_writelane_b32 v42, s29, 16
; SI-NEXT:    v_writelane_b32 v42, s21, 17
; SI-NEXT:    v_writelane_b32 v42, s85, 18
; SI-NEXT:    v_writelane_b32 v42, s84, 19
; SI-NEXT:    v_writelane_b32 v42, s13, 20
; SI-NEXT:    v_writelane_b32 v42, s65, 21
; SI-NEXT:    v_writelane_b32 v42, s98, 22
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_readfirstlane_b32 s91, v1
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_readfirstlane_b32 s16, v2
; SI-NEXT:    v_writelane_b32 v43, s16, 40
; SI-NEXT:    v_writelane_b32 v42, s14, 23
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_readfirstlane_b32 s15, v51
; SI-NEXT:    v_writelane_b32 v43, s91, 41
; SI-NEXT:    v_writelane_b32 v42, s55, 24
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_readfirstlane_b32 s19, v9
; SI-NEXT:    v_writelane_b32 v43, s15, 42
; SI-NEXT:    v_writelane_b32 v42, s82, 25
; SI-NEXT:    v_readfirstlane_b32 s20, v8
; SI-NEXT:    v_writelane_b32 v43, s19, 43
; SI-NEXT:    v_writelane_b32 v42, s23, 26
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_readfirstlane_b32 s40, v10
; SI-NEXT:    v_writelane_b32 v43, s20, 44
; SI-NEXT:    v_writelane_b32 v42, s94, 27
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_readfirstlane_b32 s41, v39
; SI-NEXT:    v_writelane_b32 v43, s40, 45
; SI-NEXT:    v_writelane_b32 v42, s60, 28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_readfirstlane_b32 s42, v49
; SI-NEXT:    v_writelane_b32 v43, s41, 46
; SI-NEXT:    v_writelane_b32 v42, s44, 29
; SI-NEXT:    v_readfirstlane_b32 s43, v48
; SI-NEXT:    v_writelane_b32 v43, s42, 47
; SI-NEXT:    v_writelane_b32 v42, s72, 30
; SI-NEXT:    v_readfirstlane_b32 s76, v13
; SI-NEXT:    v_writelane_b32 v43, s43, 48
; SI-NEXT:    v_writelane_b32 v42, s63, 31
; SI-NEXT:    v_readfirstlane_b32 s77, v12
; SI-NEXT:    v_writelane_b32 v43, s76, 49
; SI-NEXT:    v_writelane_b32 v42, s73, 32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_readfirstlane_b32 s6, v50
; SI-NEXT:    v_writelane_b32 v43, s77, 50
; SI-NEXT:    v_writelane_b32 v42, s45, 33
; SI-NEXT:    v_readfirstlane_b32 s7, v0
; SI-NEXT:    v_writelane_b32 v43, s6, 51
; SI-NEXT:    v_writelane_b32 v42, s46, 34
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v14
; SI-NEXT:    v_writelane_b32 v43, s7, 52
; SI-NEXT:    v_writelane_b32 v42, s74, 35
; SI-NEXT:    v_readfirstlane_b32 vcc_hi, v35
; SI-NEXT:    v_writelane_b32 v43, vcc_lo, 53
; SI-NEXT:    v_writelane_b32 v42, s75, 36
; SI-NEXT:    v_readfirstlane_b32 s30, v37
; SI-NEXT:    v_writelane_b32 v43, vcc_hi, 54
; SI-NEXT:    v_writelane_b32 v42, s47, 37
; SI-NEXT:    v_readfirstlane_b32 s34, v36
; SI-NEXT:    v_writelane_b32 v43, s30, 55
; SI-NEXT:    v_writelane_b32 v42, s79, 38
; SI-NEXT:    v_writelane_b32 v43, s34, 56
; SI-NEXT:    v_writelane_b32 v42, s57, 39
; SI-NEXT:    v_writelane_b32 v43, s36, 57
; SI-NEXT:    v_writelane_b32 v42, s58, 40
; SI-NEXT:    v_readfirstlane_b32 s8, v38
; SI-NEXT:    v_writelane_b32 v43, s38, 58
; SI-NEXT:    v_writelane_b32 v42, s56, 41
; SI-NEXT:    v_readfirstlane_b32 s9, v7
; SI-NEXT:    v_writelane_b32 v43, s8, 59
; SI-NEXT:    v_writelane_b32 v42, s89, 42
; SI-NEXT:    v_writelane_b32 v43, s9, 60
; SI-NEXT:    v_writelane_b32 v42, s90, 43
; SI-NEXT:    v_writelane_b32 v43, s39, 61
; SI-NEXT:    v_writelane_b32 v42, s62, 44
; SI-NEXT:    v_readfirstlane_b32 s49, v33
; SI-NEXT:    v_writelane_b32 v43, s48, 62
; SI-NEXT:    v_writelane_b32 v42, s59, 45
; SI-NEXT:    v_writelane_b32 v43, s49, 63
; SI-NEXT:    v_writelane_b32 v42, s96, 46
; SI-NEXT:    s_cbranch_scc0 .LBB97_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_readlane_b32 s4, v43, 13
; SI-NEXT:    v_readlane_b32 s5, v43, 12
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 47
; SI-NEXT:    v_readlane_b32 s4, v43, 5
; SI-NEXT:    v_readlane_b32 s5, v43, 4
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 49
; SI-NEXT:    s_and_b32 s4, s37, 0xff
; SI-NEXT:    s_lshl_b32 s5, s28, 8
; SI-NEXT:    v_writelane_b32 v42, s97, 50
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 51
; SI-NEXT:    s_and_b32 s4, s24, 0xff
; SI-NEXT:    s_lshl_b32 s5, s25, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_readlane_b32 s5, v43, 25
; SI-NEXT:    v_writelane_b32 v42, s4, 52
; SI-NEXT:    s_and_b32 s4, s71, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 53
; SI-NEXT:    v_readlane_b32 s4, v43, 18
; SI-NEXT:    v_readlane_b32 s5, v43, 17
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 54
; SI-NEXT:    s_and_b32 s4, s7, 0xff
; SI-NEXT:    s_lshl_b32 s5, s6, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 55
; SI-NEXT:    s_and_b32 s4, s9, 0xff
; SI-NEXT:    s_lshl_b32 s5, s8, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 56
; SI-NEXT:    s_and_b32 s4, s11, 0xff
; SI-NEXT:    s_lshl_b32 s5, s10, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 57
; SI-NEXT:    s_and_b32 s4, s12, 0xff
; SI-NEXT:    s_lshl_b32 s5, s18, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    v_writelane_b32 v42, s4, 58
; SI-NEXT:    s_and_b32 s4, s13, 0xff
; SI-NEXT:    s_lshl_b32 s5, s17, 8
; SI-NEXT:    s_or_b32 s97, s4, s5
; SI-NEXT:    s_and_b32 s4, s74, 0xff
; SI-NEXT:    s_lshl_b32 s5, s63, 8
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s73, 0xff
; SI-NEXT:    s_lshl_b32 s5, s72, 8
; SI-NEXT:    s_or_b32 s72, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xff
; SI-NEXT:    s_lshl_b32 s5, s14, 8
; SI-NEXT:    s_or_b32 s73, s4, s5
; SI-NEXT:    s_and_b32 s4, s79, 0xff
; SI-NEXT:    s_lshl_b32 s5, s75, 8
; SI-NEXT:    s_or_b32 s74, s4, s5
; SI-NEXT:    s_and_b32 s4, s90, 0xff
; SI-NEXT:    s_lshl_b32 s5, s89, 8
; SI-NEXT:    s_or_b32 s75, s4, s5
; SI-NEXT:    v_readlane_b32 s4, v43, 9
; SI-NEXT:    v_readlane_b32 s5, v43, 8
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s5, s4, s5
; SI-NEXT:    v_readlane_b32 s4, v43, 7
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    v_readlane_b32 s6, v43, 6
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_or_b32 s79, s6, s4
; SI-NEXT:    v_readlane_b32 s4, v43, 11
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    v_readlane_b32 s6, v43, 10
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_or_b32 s4, s6, s4
; SI-NEXT:    v_readlane_b32 s6, v43, 1
; SI-NEXT:    v_readlane_b32 s7, v43, 0
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_or_b32 s7, s6, s7
; SI-NEXT:    v_readlane_b32 s6, v43, 38
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s8, v43, 37
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_or_b32 s89, s8, s6
; SI-NEXT:    v_readlane_b32 s6, v43, 3
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s8, v43, 2
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_or_b32 s6, s8, s6
; SI-NEXT:    v_readlane_b32 s8, v43, 34
; SI-NEXT:    v_readlane_b32 s9, v43, 33
; SI-NEXT:    v_writelane_b32 v42, s26, 59
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s9, s9, 8
; SI-NEXT:    v_writelane_b32 v42, s31, 60
; SI-NEXT:    s_or_b32 s9, s8, s9
; SI-NEXT:    v_readlane_b32 s8, v43, 32
; SI-NEXT:    v_writelane_b32 v42, s17, 61
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v43, 31
; SI-NEXT:    v_writelane_b32 v42, s35, 62
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    v_writelane_b32 v42, s27, 63
; SI-NEXT:    s_or_b32 s27, s10, s8
; SI-NEXT:    v_readlane_b32 s8, v43, 36
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v43, 35
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_or_b32 s8, s10, s8
; SI-NEXT:    v_readlane_b32 s10, v43, 29
; SI-NEXT:    v_readlane_b32 s11, v43, 28
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    s_lshl_b32 s11, s11, 8
; SI-NEXT:    s_or_b32 s11, s10, s11
; SI-NEXT:    v_readlane_b32 s10, v43, 27
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s12, v43, 26
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s12, s12, 24
; SI-NEXT:    s_or_b32 s93, s12, s10
; SI-NEXT:    v_readlane_b32 s10, v43, 30
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s12, s70, 24
; SI-NEXT:    s_or_b32 s10, s12, s10
; SI-NEXT:    v_readlane_b32 s12, v43, 22
; SI-NEXT:    v_readlane_b32 s13, v43, 21
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    s_lshl_b32 s13, s13, 8
; SI-NEXT:    s_or_b32 s13, s12, s13
; SI-NEXT:    v_readlane_b32 s12, v43, 20
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s14, v43, 19
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_or_b32 s26, s14, s12
; SI-NEXT:    v_readlane_b32 s12, v43, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s14, v43, 23
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_or_b32 s12, s14, s12
; SI-NEXT:    v_readlane_b32 s14, v43, 14
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s15, s15, 8
; SI-NEXT:    s_or_b32 s15, s14, s15
; SI-NEXT:    s_and_b32 s14, s16, 0xff
; SI-NEXT:    ; implicit-def: $vgpr41 : SGPR spill to VGPR lane
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    v_writelane_b32 v41, s18, 0
; SI-NEXT:    s_lshl_b32 s18, s91, 24
; SI-NEXT:    s_or_b32 s68, s18, s14
; SI-NEXT:    v_readlane_b32 s14, v43, 16
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s16, v43, 15
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_lshl_b32 s18, s16, 24
; SI-NEXT:    s_or_b32 s14, s18, s14
; SI-NEXT:    s_and_b32 s18, s41, 0xff
; SI-NEXT:    s_lshl_b32 s40, s40, 8
; SI-NEXT:    s_or_b32 s41, s18, s40
; SI-NEXT:    s_and_b32 s18, s19, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_lshl_b32 s40, s20, 24
; SI-NEXT:    s_mov_b32 s95, s24
; SI-NEXT:    s_or_b32 s24, s40, s18
; SI-NEXT:    s_and_b32 s40, s42, 0xff
; SI-NEXT:    s_lshl_b32 s40, s40, 16
; SI-NEXT:    s_lshl_b32 s42, s43, 24
; SI-NEXT:    s_or_b32 s40, s42, s40
; SI-NEXT:    s_and_b32 s42, vcc_hi, 0xff
; SI-NEXT:    s_lshl_b32 s43, vcc_lo, 8
; SI-NEXT:    s_or_b32 s43, s42, s43
; SI-NEXT:    s_and_b32 s42, s76, 0xff
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s76, s77, 24
; SI-NEXT:    s_mov_b32 s35, s69
; SI-NEXT:    s_or_b32 s69, s76, s42
; SI-NEXT:    s_and_b32 s42, s30, 0xff
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s76, s34, 24
; SI-NEXT:    s_or_b32 s42, s76, s42
; SI-NEXT:    s_and_b32 s76, s48, 0xff
; SI-NEXT:    s_lshl_b32 s77, s39, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s36, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s38, 24
; SI-NEXT:    v_writelane_b32 v41, s70, 1
; SI-NEXT:    s_or_b32 s70, s78, s77
; SI-NEXT:    s_and_b32 s77, s49, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s50, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 vcc_lo, s78, s77
; SI-NEXT:    s_or_b32 vcc_hi, s76, s70
; SI-NEXT:    s_and_b32 s76, s54, 0xff
; SI-NEXT:    s_lshl_b32 s77, s53, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s51, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s52, 24
; SI-NEXT:    v_writelane_b32 v41, s71, 2
; SI-NEXT:    s_or_b32 s71, s78, s77
; SI-NEXT:    s_and_b32 s77, s64, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s66, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s38, s78, s77
; SI-NEXT:    s_or_b32 s39, s76, s71
; SI-NEXT:    s_and_b32 s76, s29, 0xff
; SI-NEXT:    s_lshl_b32 s77, s81, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s67, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s61, 24
; SI-NEXT:    v_readlane_b32 s16, v42, 63
; SI-NEXT:    s_mov_b32 s99, s28
; SI-NEXT:    s_mov_b32 s28, s80
; SI-NEXT:    s_or_b32 s80, s78, s77
; SI-NEXT:    s_and_b32 s77, s16, 0xff
; SI-NEXT:    v_readlane_b32 s16, v42, 62
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s48, s78, s77
; SI-NEXT:    s_or_b32 s49, s76, s80
; SI-NEXT:    s_and_b32 s76, s35, 0xff
; SI-NEXT:    s_lshl_b32 s77, s85, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s87, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s86, 24
; SI-NEXT:    s_or_b32 s81, s78, s77
; SI-NEXT:    s_and_b32 s77, s94, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s82, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s50, s78, s77
; SI-NEXT:    s_or_b32 s51, s76, s81
; SI-NEXT:    s_and_b32 s76, s22, 0xff
; SI-NEXT:    s_lshl_b32 s77, s21, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s83, 0xff
; SI-NEXT:    v_readlane_b32 s16, v43, 39
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_or_b32 s82, s78, s77
; SI-NEXT:    s_and_b32 s77, s55, 0xff
; SI-NEXT:    v_readlane_b32 s16, v42, 50
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s52, s78, s77
; SI-NEXT:    s_or_b32 s53, s76, s82
; SI-NEXT:    s_and_b32 s76, s92, 0xff
; SI-NEXT:    s_lshl_b32 s77, s28, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s84, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s96, 24
; SI-NEXT:    v_readlane_b32 s16, v42, 59
; SI-NEXT:    s_or_b32 s20, s78, s77
; SI-NEXT:    s_and_b32 s77, s16, 0xff
; SI-NEXT:    v_readlane_b32 s16, v42, 60
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s16, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_or_b32 s54, s78, s77
; SI-NEXT:    s_or_b32 s55, s76, s20
; SI-NEXT:    s_and_b32 s76, s44, 0xff
; SI-NEXT:    s_lshl_b32 s77, s60, 8
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s65, 0xff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s98, 24
; SI-NEXT:    v_readlane_b32 s16, v42, 47
; SI-NEXT:    s_or_b32 s21, s78, s77
; SI-NEXT:    s_and_b32 s77, s45, 0xff
; SI-NEXT:    s_and_b32 s44, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 49
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s46, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_and_b32 s45, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 51
; SI-NEXT:    s_or_b32 s64, s78, s77
; SI-NEXT:    s_or_b32 s65, s76, s21
; SI-NEXT:    s_and_b32 s76, s58, 0xff
; SI-NEXT:    s_lshl_b32 s77, s57, 8
; SI-NEXT:    s_and_b32 s46, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 52
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_and_b32 s77, s47, 0xff
; SI-NEXT:    s_and_b32 s47, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 53
; SI-NEXT:    s_lshl_b32 s78, s56, 24
; SI-NEXT:    s_and_b32 s56, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 54
; SI-NEXT:    s_and_b32 s57, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 55
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_and_b32 s30, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 56
; SI-NEXT:    s_or_b32 s7, s7, s89
; SI-NEXT:    s_or_b32 s96, s78, s77
; SI-NEXT:    s_and_b32 s77, s62, 0xff
; SI-NEXT:    s_and_b32 s34, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 57
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s77, s77, 16
; SI-NEXT:    s_lshl_b32 s78, s59, 24
; SI-NEXT:    s_and_b32 s76, s76, 0xffff
; SI-NEXT:    s_and_b32 s36, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s16, v42, 58
; SI-NEXT:    s_and_b32 s86, s97, 0xffff
; SI-NEXT:    s_and_b32 s87, s72, 0xffff
; SI-NEXT:    s_and_b32 s97, s73, 0xffff
; SI-NEXT:    s_or_b32 s72, s45, s6
; SI-NEXT:    s_mov_b32 s73, s7
; SI-NEXT:    s_lshr_b64 s[18:19], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[6:7], vcc, 16
; SI-NEXT:    s_or_b32 s5, s5, s79
; SI-NEXT:    s_or_b32 s9, s9, s27
; SI-NEXT:    s_or_b32 s11, s11, s93
; SI-NEXT:    s_or_b32 s13, s13, s26
; SI-NEXT:    s_or_b32 s15, s15, s68
; SI-NEXT:    s_or_b32 s41, s41, s24
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_mov_b32 s29, s99
; SI-NEXT:    s_or_b32 s66, s78, s77
; SI-NEXT:    s_or_b32 s67, s76, s96
; SI-NEXT:    s_and_b32 s83, s16, 0xffff
; SI-NEXT:    s_and_b32 s98, s63, 0xffff
; SI-NEXT:    s_and_b32 s99, s74, 0xffff
; SI-NEXT:    s_and_b32 s78, s75, 0xffff
; SI-NEXT:    v_writelane_b32 v42, s6, 47
; SI-NEXT:    s_mov_b32 s88, s22
; SI-NEXT:    s_or_b32 s43, s43, s69
; SI-NEXT:    s_mov_b32 s31, s95
; SI-NEXT:    s_or_b32 s74, s44, s4
; SI-NEXT:    s_mov_b32 s75, s5
; SI-NEXT:    s_lshr_b64 s[22:23], s[4:5], 16
; SI-NEXT:    s_or_b32 s62, s46, s8
; SI-NEXT:    s_lshr_b64 s[94:95], s[8:9], 16
; SI-NEXT:    s_or_b32 s60, s47, s10
; SI-NEXT:    s_mov_b32 s61, s11
; SI-NEXT:    s_lshr_b64 s[90:91], s[10:11], 16
; SI-NEXT:    s_or_b32 s58, s56, s12
; SI-NEXT:    s_mov_b32 s59, s13
; SI-NEXT:    s_lshr_b64 s[4:5], s[12:13], 16
; SI-NEXT:    s_or_b32 s56, s57, s14
; SI-NEXT:    s_mov_b32 s57, s15
; SI-NEXT:    s_lshr_b64 s[76:77], s[14:15], 16
; SI-NEXT:    s_or_b32 s46, s30, s40
; SI-NEXT:    s_mov_b32 s47, s41
; SI-NEXT:    s_lshr_b64 s[16:17], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v42, s7, 48
; SI-NEXT:    s_or_b32 s40, s83, s38
; SI-NEXT:    s_mov_b32 s41, s39
; SI-NEXT:    s_lshr_b64 s[38:39], s[38:39], 16
; SI-NEXT:    s_or_b32 s14, s86, s48
; SI-NEXT:    s_mov_b32 s15, s49
; SI-NEXT:    s_lshr_b64 s[48:49], s[48:49], 16
; SI-NEXT:    s_or_b32 s12, s98, s50
; SI-NEXT:    s_mov_b32 s13, s51
; SI-NEXT:    s_lshr_b64 s[50:51], s[50:51], 16
; SI-NEXT:    s_or_b32 s10, s87, s52
; SI-NEXT:    s_mov_b32 s11, s53
; SI-NEXT:    s_lshr_b64 s[52:53], s[52:53], 16
; SI-NEXT:    s_or_b32 s6, s99, s64
; SI-NEXT:    s_or_b32 s98, s78, s66
; SI-NEXT:    s_mov_b32 s99, s67
; SI-NEXT:    s_lshr_b64 s[66:67], s[66:67], 16
; SI-NEXT:    s_mov_b32 s63, s9
; SI-NEXT:    s_or_b32 s44, s34, s42
; SI-NEXT:    s_mov_b32 s45, s43
; SI-NEXT:    s_lshr_b64 s[84:85], s[42:43], 16
; SI-NEXT:    s_or_b32 s42, s36, vcc_lo
; SI-NEXT:    s_mov_b32 s43, vcc_hi
; SI-NEXT:    s_or_b32 s8, s97, s54
; SI-NEXT:    s_mov_b32 s9, s55
; SI-NEXT:    s_lshr_b64 s[54:55], s[54:55], 16
; SI-NEXT:    v_readlane_b32 s97, v42, 50
; SI-NEXT:    s_mov_b32 s7, s65
; SI-NEXT:    s_lshr_b64 s[64:65], s[64:65], 16
; SI-NEXT:    s_lshr_b32 s19, s79, 16
; SI-NEXT:    s_lshr_b32 s67, s89, 16
; SI-NEXT:    s_lshr_b32 s5, s27, 16
; SI-NEXT:    s_lshr_b32 s79, s93, 16
; SI-NEXT:    s_lshr_b32 s89, s26, 16
; SI-NEXT:    s_lshr_b32 s91, s68, 16
; SI-NEXT:    s_lshr_b32 s93, s24, 16
; SI-NEXT:    s_mov_b32 s86, s18
; SI-NEXT:    s_lshr_b32 s95, s69, 16
; SI-NEXT:    s_lshr_b32 s30, s70, 16
; SI-NEXT:    v_readlane_b32 s70, v41, 1
; SI-NEXT:    s_lshr_b32 s34, s71, 16
; SI-NEXT:    s_mov_b32 s24, s31
; SI-NEXT:    v_readlane_b32 s71, v41, 2
; SI-NEXT:    s_lshr_b32 s36, s80, 16
; SI-NEXT:    s_lshr_b32 s39, s81, 16
; SI-NEXT:    s_lshr_b32 s49, s82, 16
; SI-NEXT:    s_mov_b32 s82, s4
; SI-NEXT:    s_mov_b32 s4, s22
; SI-NEXT:    s_lshr_b32 s51, s20, 16
; SI-NEXT:    v_readlane_b32 s18, v41, 0
; SI-NEXT:    s_lshr_b32 s53, s21, 16
; SI-NEXT:    s_mov_b32 s69, s35
; SI-NEXT:    v_readlane_b32 s27, v42, 63
; SI-NEXT:    v_readlane_b32 s35, v42, 62
; SI-NEXT:    v_readlane_b32 s17, v42, 61
; SI-NEXT:    v_readlane_b32 s31, v42, 60
; SI-NEXT:    v_readlane_b32 s26, v42, 59
; SI-NEXT:    s_mov_b32 s80, s28
; SI-NEXT:    s_mov_b32 s22, s88
; SI-NEXT:    s_lshr_b32 s68, s96, 16
; SI-NEXT:    s_mov_b32 s28, s29
; SI-NEXT:    s_cbranch_execnz .LBB97_3
; SI-NEXT:  .LBB97_2: ; %cmp.true
; SI-NEXT:    v_readlane_b32 s4, v42, 43
; SI-NEXT:    s_add_i32 s19, s4, 3
; SI-NEXT:    v_readlane_b32 s5, v42, 42
; SI-NEXT:    v_readlane_b32 s6, v42, 44
; SI-NEXT:    s_and_b32 s4, s19, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_add_i32 s20, s6, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readlane_b32 s5, v42, 45
; SI-NEXT:    s_and_b32 s6, s20, 0xff
; SI-NEXT:    s_lshl_b32 s5, s5, 24
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_addk_i32 s4, 0x300
; SI-NEXT:    s_or_b32 s5, s5, s6
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readlane_b32 s5, v42, 40
; SI-NEXT:    s_add_i32 s23, s5, 3
; SI-NEXT:    v_readlane_b32 s6, v42, 39
; SI-NEXT:    v_readlane_b32 s7, v42, 37
; SI-NEXT:    s_and_b32 s5, s23, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_readlane_b32 s6, v42, 41
; SI-NEXT:    s_and_b32 s7, s7, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_addk_i32 s5, 0x300
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_readlane_b32 s6, v42, 38
; SI-NEXT:    s_add_i32 s68, s6, 3
; SI-NEXT:    v_readlane_b32 s7, v42, 36
; SI-NEXT:    v_readlane_b32 s8, v42, 33
; SI-NEXT:    s_and_b32 s6, s68, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_add_i32 s87, s8, 3
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readlane_b32 s7, v42, 34
; SI-NEXT:    s_and_b32 s8, s87, 0xff
; SI-NEXT:    s_lshl_b32 s7, s7, 24
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_addk_i32 s6, 0x300
; SI-NEXT:    s_or_b32 s7, s7, s8
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readlane_b32 s7, v42, 29
; SI-NEXT:    s_add_i32 s83, s7, 3
; SI-NEXT:    v_readlane_b32 s8, v42, 28
; SI-NEXT:    v_readlane_b32 s9, v42, 21
; SI-NEXT:    s_and_b32 s7, s83, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_readlane_b32 s8, v42, 22
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_addk_i32 s7, 0x300
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_readlane_b32 s8, v42, 26
; SI-NEXT:    s_add_i32 s8, s8, 3
; SI-NEXT:    v_readlane_b32 s9, v42, 23
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    s_lshl_b32 s9, s9, 8
; SI-NEXT:    s_add_i32 s10, s26, 3
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    s_lshl_b32 s9, s31, 24
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_addk_i32 s8, 0x300
; SI-NEXT:    s_or_b32 s9, s9, s10
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_add_i32 s9, s92, 3
; SI-NEXT:    v_readlane_b32 s11, v42, 19
; SI-NEXT:    s_and_b32 s9, s9, 0xff
; SI-NEXT:    s_lshl_b32 s10, s80, 8
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_add_i32 s98, s4, 0x3000000
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_readlane_b32 s4, v42, 46
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s10, s4, 24
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_addk_i32 s9, 0x300
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_readlane_b32 s10, v42, 32
; SI-NEXT:    s_add_i32 s86, s10, 3
; SI-NEXT:    v_readlane_b32 s11, v42, 30
; SI-NEXT:    v_readlane_b32 s12, v42, 24
; SI-NEXT:    s_and_b32 s10, s86, 0xff
; SI-NEXT:    s_lshl_b32 s11, s11, 8
; SI-NEXT:    s_add_i32 s12, s12, 3
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    s_lshl_b32 s11, s97, 24
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_addk_i32 s10, 0x300
; SI-NEXT:    s_or_b32 s11, s11, s12
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    s_add_i32 s11, s22, 3
; SI-NEXT:    v_readlane_b32 s12, v42, 17
; SI-NEXT:    v_readlane_b32 s13, v42, 11
; SI-NEXT:    s_and_b32 s11, s11, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_readlane_b32 s12, v43, 39
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s12, s12, 24
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_addk_i32 s11, 0x300
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_readlane_b32 s12, v42, 35
; SI-NEXT:    s_add_i32 s97, s12, 3
; SI-NEXT:    v_readlane_b32 s13, v42, 31
; SI-NEXT:    v_readlane_b32 s14, v42, 27
; SI-NEXT:    s_and_b32 s12, s97, 0xff
; SI-NEXT:    s_lshl_b32 s13, s13, 8
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    v_readlane_b32 s13, v42, 25
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s13, s13, 24
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_addk_i32 s12, 0x300
; SI-NEXT:    s_or_b32 s13, s13, s14
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    s_add_i32 s13, s69, 3
; SI-NEXT:    v_readlane_b32 s14, v42, 18
; SI-NEXT:    v_readlane_b32 s15, v42, 14
; SI-NEXT:    s_and_b32 s13, s13, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 8
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_readlane_b32 s14, v42, 12
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_addk_i32 s13, 0x300
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_readlane_b32 s14, v42, 20
; SI-NEXT:    s_add_i32 s14, s14, 3
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    s_lshl_b32 s15, s17, 8
; SI-NEXT:    s_add_i32 s16, s27, 3
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s15, s35, 24
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_addk_i32 s14, 0x300
; SI-NEXT:    s_or_b32 s15, s15, s16
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    v_readlane_b32 s15, v42, 16
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    v_readlane_b32 s16, v42, 15
; SI-NEXT:    v_readlane_b32 s17, v42, 9
; SI-NEXT:    s_and_b32 s15, s15, 0xff
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    v_readlane_b32 s16, v42, 10
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s16, s16, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_addk_i32 s15, 0x300
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    v_readlane_b32 s16, v42, 13
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_lshl_b32 s17, s18, 8
; SI-NEXT:    v_readlane_b32 s18, v42, 7
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 8
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 6
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    v_readlane_b32 s18, v42, 5
; SI-NEXT:    v_readlane_b32 s19, v42, 2
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 8
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_readlane_b32 s18, v42, 3
; SI-NEXT:    s_and_b32 s19, s19, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_add_i32 s40, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v42, 4
; SI-NEXT:    s_add_i32 s41, s17, 0x3000000
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v42, 1
; SI-NEXT:    v_readlane_b32 s18, v43, 63
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v42, 0
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 62
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    v_readlane_b32 s18, v43, 61
; SI-NEXT:    s_and_b32 s17, s17, 0xff
; SI-NEXT:    s_lshl_b32 s18, s18, 8
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_readlane_b32 s18, v43, 57
; SI-NEXT:    s_addk_i32 s17, 0x300
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s42, s16, 0x3000000
; SI-NEXT:    s_and_b32 s16, s17, 0xffff
; SI-NEXT:    v_readlane_b32 s17, v43, 58
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s43, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 60
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 59
; SI-NEXT:    v_readlane_b32 s18, v43, 55
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 56
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s44, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 54
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 53
; SI-NEXT:    v_readlane_b32 s18, v43, 49
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 50
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s45, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 52
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 51
; SI-NEXT:    v_readlane_b32 s18, v43, 47
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 48
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s46, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 46
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 45
; SI-NEXT:    v_readlane_b32 s18, v43, 43
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 44
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s47, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 18
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 17
; SI-NEXT:    v_readlane_b32 s18, v43, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 15
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s56, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 14
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 42
; SI-NEXT:    v_readlane_b32 s18, v43, 40
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 41
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s57, s16, 0x3000000
; SI-NEXT:    s_add_i32 s16, s71, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 25
; SI-NEXT:    v_readlane_b32 s18, v43, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 23
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s58, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 22
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 21
; SI-NEXT:    v_readlane_b32 s18, v43, 20
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 19
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s59, s16, 0x3000000
; SI-NEXT:    s_add_i32 s16, s24, 3
; SI-NEXT:    v_readlane_b32 s18, v43, 30
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s25, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s70, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s60, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 29
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 28
; SI-NEXT:    v_readlane_b32 s18, v43, 27
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 26
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s61, s16, 0x3000000
; SI-NEXT:    s_add_i32 s16, s37, 3
; SI-NEXT:    v_readlane_b32 s18, v43, 36
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s28, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 35
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s62, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 34
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 33
; SI-NEXT:    v_readlane_b32 s18, v43, 32
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 31
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s63, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 5
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 4
; SI-NEXT:    v_readlane_b32 s18, v43, 3
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 2
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s72, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 1
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 0
; SI-NEXT:    v_readlane_b32 s18, v43, 38
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 37
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s73, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 13
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 12
; SI-NEXT:    v_readlane_b32 s18, v43, 11
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 10
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s74, s16, 0x3000000
; SI-NEXT:    v_readlane_b32 s16, v43, 9
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v43, 8
; SI-NEXT:    v_readlane_b32 s18, v43, 7
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_readlane_b32 s17, v43, 6
; SI-NEXT:    s_and_b32 s18, s18, 0xff
; SI-NEXT:    s_addk_i32 s16, 0x300
; SI-NEXT:    s_lshl_b32 s17, s17, 24
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s99, s5, 0x3000000
; SI-NEXT:    s_add_i32 s10, s10, 0x3000000
; SI-NEXT:    s_add_i32 s11, s11, 0x3000000
; SI-NEXT:    s_add_i32 s12, s12, 0x3000000
; SI-NEXT:    s_add_i32 s13, s13, 0x3000000
; SI-NEXT:    s_add_i32 s14, s14, 0x3000000
; SI-NEXT:    s_add_i32 s15, s15, 0x3000000
; SI-NEXT:    s_add_i32 s75, s16, 0x3000000
; SI-NEXT:    s_lshr_b64 s[18:19], s[42:43], 16
; SI-NEXT:    s_add_i32 s6, s6, 0x3000000
; SI-NEXT:    s_add_i32 s7, s7, 0x3000000
; SI-NEXT:    s_add_i32 s8, s8, 0x3000000
; SI-NEXT:    s_add_i32 s9, s9, 0x3000000
; SI-NEXT:    s_lshr_b64 s[4:5], s[74:75], 16
; SI-NEXT:    s_lshr_b64 s[94:95], s[62:63], 16
; SI-NEXT:    s_lshr_b64 s[90:91], s[60:61], 16
; SI-NEXT:    v_writelane_b32 v42, s18, 47
; SI-NEXT:    s_lshr_b64 s[38:39], s[40:41], 16
; SI-NEXT:    s_lshr_b64 s[48:49], s[14:15], 16
; SI-NEXT:    s_lshr_b64 s[50:51], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[52:53], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[66:67], s[98:99], 16
; SI-NEXT:    s_lshr_b64 s[86:87], s[72:73], 16
; SI-NEXT:    s_lshr_b64 s[82:83], s[58:59], 16
; SI-NEXT:    s_lshr_b64 s[76:77], s[56:57], 16
; SI-NEXT:    s_lshr_b64 s[16:17], s[46:47], 16
; SI-NEXT:    s_lshr_b64 s[84:85], s[44:45], 16
; SI-NEXT:    v_writelane_b32 v42, s19, 48
; SI-NEXT:    s_lshr_b64 s[54:55], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[64:65], s[6:7], 16
; SI-NEXT:    s_lshr_b32 s19, s75, 16
; SI-NEXT:    s_lshr_b32 s67, s73, 16
; SI-NEXT:    s_lshr_b32 s5, s63, 16
; SI-NEXT:    s_lshr_b32 s79, s61, 16
; SI-NEXT:    s_lshr_b32 s89, s59, 16
; SI-NEXT:    s_lshr_b32 s91, s57, 16
; SI-NEXT:    s_lshr_b32 s93, s47, 16
; SI-NEXT:    s_lshr_b32 s95, s45, 16
; SI-NEXT:    s_lshr_b32 s30, s43, 16
; SI-NEXT:    s_lshr_b32 s34, s41, 16
; SI-NEXT:    s_lshr_b32 s36, s15, 16
; SI-NEXT:    s_lshr_b32 s39, s13, 16
; SI-NEXT:    s_lshr_b32 s49, s11, 16
; SI-NEXT:    s_lshr_b32 s51, s9, 16
; SI-NEXT:    s_lshr_b32 s53, s7, 16
; SI-NEXT:    s_lshr_b32 s68, s99, 16
; SI-NEXT:  .LBB97_3: ; %end
; SI-NEXT:    s_and_b32 s17, s74, 0xffff
; SI-NEXT:    s_lshl_b32 s18, s4, 16
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_and_b32 s18, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_and_b32 s19, s72, 0xffff
; SI-NEXT:    s_lshl_b32 s20, s86, 16
; SI-NEXT:    s_or_b32 s19, s19, s20
; SI-NEXT:    s_and_b32 s20, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s21, s67, 16
; SI-NEXT:    s_or_b32 s20, s20, s21
; SI-NEXT:    s_and_b32 s21, s62, 0xffff
; SI-NEXT:    s_lshl_b32 s22, s94, 16
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_and_b32 s22, s63, 0xffff
; SI-NEXT:    s_lshl_b32 s23, s5, 16
; SI-NEXT:    s_or_b32 s22, s22, s23
; SI-NEXT:    s_and_b32 s23, s60, 0xffff
; SI-NEXT:    s_lshl_b32 s24, s90, 16
; SI-NEXT:    s_or_b32 s23, s23, s24
; SI-NEXT:    s_and_b32 s24, s61, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s79, 16
; SI-NEXT:    s_or_b32 s24, s24, s25
; SI-NEXT:    s_and_b32 s25, s58, 0xffff
; SI-NEXT:    s_lshl_b32 s26, s82, 16
; SI-NEXT:    s_or_b32 s25, s25, s26
; SI-NEXT:    s_and_b32 s26, s59, 0xffff
; SI-NEXT:    s_lshl_b32 s27, s89, 16
; SI-NEXT:    s_or_b32 s26, s26, s27
; SI-NEXT:    s_and_b32 s27, s56, 0xffff
; SI-NEXT:    s_lshl_b32 s28, s76, 16
; SI-NEXT:    s_or_b32 s27, s27, s28
; SI-NEXT:    s_and_b32 s28, s57, 0xffff
; SI-NEXT:    s_lshl_b32 s29, s91, 16
; SI-NEXT:    s_or_b32 s28, s28, s29
; SI-NEXT:    s_and_b32 s29, s46, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s16, 16
; SI-NEXT:    s_or_b32 s29, s29, s46
; SI-NEXT:    s_and_b32 s46, s47, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s93, 16
; SI-NEXT:    s_or_b32 s46, s46, s47
; SI-NEXT:    s_and_b32 s44, s44, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s84, 16
; SI-NEXT:    s_or_b32 s44, s44, s47
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s95, 16
; SI-NEXT:    v_readlane_b32 s4, v42, 47
; SI-NEXT:    s_or_b32 s45, s45, s47
; SI-NEXT:    s_and_b32 s42, s42, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s4, 16
; SI-NEXT:    s_or_b32 s42, s42, s47
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s30, 16
; SI-NEXT:    s_or_b32 s43, s43, s47
; SI-NEXT:    s_and_b32 s40, s40, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s38, 16
; SI-NEXT:    s_or_b32 s40, s40, s47
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s34, 16
; SI-NEXT:    s_or_b32 s41, s41, s47
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s48, 16
; SI-NEXT:    s_or_b32 s14, s14, s47
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s36, 16
; SI-NEXT:    s_or_b32 s15, s15, s47
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s50, 16
; SI-NEXT:    s_or_b32 s12, s12, s47
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s39, 16
; SI-NEXT:    s_or_b32 s13, s13, s47
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s52, 16
; SI-NEXT:    s_or_b32 s10, s10, s47
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s49, 16
; SI-NEXT:    s_or_b32 s11, s11, s47
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s54, 16
; SI-NEXT:    s_or_b32 s8, s8, s47
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s51, 16
; SI-NEXT:    s_or_b32 s9, s9, s47
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s64, 16
; SI-NEXT:    v_readlane_b32 s5, v42, 48
; SI-NEXT:    s_or_b32 s6, s6, s47
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s53, 16
; SI-NEXT:    s_or_b32 s7, s7, s47
; SI-NEXT:    s_and_b32 s4, s98, 0xffff
; SI-NEXT:    s_lshl_b32 s47, s66, 16
; SI-NEXT:    s_and_b32 s5, s99, 0xffff
; SI-NEXT:    s_lshl_b32 s16, s68, 16
; SI-NEXT:    s_or_b32 s4, s4, s47
; SI-NEXT:    s_or_b32 s5, s5, s16
; SI-NEXT:    v_mov_b32_e32 v0, s17
; SI-NEXT:    v_mov_b32_e32 v1, s18
; SI-NEXT:    v_mov_b32_e32 v2, s19
; SI-NEXT:    v_mov_b32_e32 v3, s20
; SI-NEXT:    v_mov_b32_e32 v4, s21
; SI-NEXT:    v_mov_b32_e32 v5, s22
; SI-NEXT:    v_mov_b32_e32 v6, s23
; SI-NEXT:    v_mov_b32_e32 v7, s24
; SI-NEXT:    v_mov_b32_e32 v8, s25
; SI-NEXT:    v_mov_b32_e32 v9, s26
; SI-NEXT:    v_mov_b32_e32 v10, s27
; SI-NEXT:    v_mov_b32_e32 v11, s28
; SI-NEXT:    v_mov_b32_e32 v12, s29
; SI-NEXT:    v_mov_b32_e32 v13, s46
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_readlane_b32 s99, v40, 35
; SI-NEXT:    v_readlane_b32 s98, v40, 34
; SI-NEXT:    v_readlane_b32 s97, v40, 33
; SI-NEXT:    v_readlane_b32 s96, v40, 32
; SI-NEXT:    v_readlane_b32 s87, v40, 31
; SI-NEXT:    v_readlane_b32 s86, v40, 30
; SI-NEXT:    v_readlane_b32 s85, v40, 29
; SI-NEXT:    v_readlane_b32 s84, v40, 28
; SI-NEXT:    v_readlane_b32 s83, v40, 27
; SI-NEXT:    v_readlane_b32 s82, v40, 26
; SI-NEXT:    v_readlane_b32 s81, v40, 25
; SI-NEXT:    v_readlane_b32 s80, v40, 24
; SI-NEXT:    v_readlane_b32 s71, v40, 23
; SI-NEXT:    v_readlane_b32 s70, v40, 22
; SI-NEXT:    v_readlane_b32 s69, v40, 21
; SI-NEXT:    v_readlane_b32 s68, v40, 20
; SI-NEXT:    v_readlane_b32 s67, v40, 19
; SI-NEXT:    v_readlane_b32 s66, v40, 18
; SI-NEXT:    v_readlane_b32 s65, v40, 17
; SI-NEXT:    v_readlane_b32 s64, v40, 16
; SI-NEXT:    v_readlane_b32 s55, v40, 15
; SI-NEXT:    v_readlane_b32 s54, v40, 14
; SI-NEXT:    v_readlane_b32 s53, v40, 13
; SI-NEXT:    v_readlane_b32 s52, v40, 12
; SI-NEXT:    v_readlane_b32 s51, v40, 11
; SI-NEXT:    v_readlane_b32 s50, v40, 10
; SI-NEXT:    v_readlane_b32 s49, v40, 9
; SI-NEXT:    v_readlane_b32 s48, v40, 8
; SI-NEXT:    v_readlane_b32 s39, v40, 7
; SI-NEXT:    v_readlane_b32 s38, v40, 6
; SI-NEXT:    v_readlane_b32 s37, v40, 5
; SI-NEXT:    v_readlane_b32 s36, v40, 4
; SI-NEXT:    v_readlane_b32 s35, v40, 3
; SI-NEXT:    v_readlane_b32 s34, v40, 2
; SI-NEXT:    v_readlane_b32 s31, v40, 1
; SI-NEXT:    v_readlane_b32 s30, v40, 0
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB97_4:
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    v_writelane_b32 v42, s6, 47
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr19
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr5
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr79
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr89
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr76
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr16
; SI-NEXT:    ; implicit-def: $sgpr93
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    v_writelane_b32 v42, s7, 48
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    s_branch .LBB97_2
;
; VI-LABEL: bitcast_v128i8_to_v64i16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    v_readfirstlane_b32 s63, v12
; VI-NEXT:    v_readfirstlane_b32 s8, v0
; VI-NEXT:    buffer_load_ushort v12, off, s[0:3], s32 offset:280
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:276
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; VI-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:12
; VI-NEXT:    v_readfirstlane_b32 s59, v29
; VI-NEXT:    v_readfirstlane_b32 s47, v27
; VI-NEXT:    v_readfirstlane_b32 s75, v30
; VI-NEXT:    v_readfirstlane_b32 s72, v28
; VI-NEXT:    v_readfirstlane_b32 s60, v26
; VI-NEXT:    v_readfirstlane_b32 s43, v25
; VI-NEXT:    v_readfirstlane_b32 s56, v24
; VI-NEXT:    v_readfirstlane_b32 s15, v23
; VI-NEXT:    v_readfirstlane_b32 s44, v22
; VI-NEXT:    v_readfirstlane_b32 s12, v21
; VI-NEXT:    v_readfirstlane_b32 s40, v20
; VI-NEXT:    v_readfirstlane_b32 s76, v18
; VI-NEXT:    v_readfirstlane_b32 s73, v16
; VI-NEXT:    v_readfirstlane_b32 s9, v19
; VI-NEXT:    v_readfirstlane_b32 s61, v17
; VI-NEXT:    v_readfirstlane_b32 s62, v15
; VI-NEXT:    v_readfirstlane_b32 s74, v14
; VI-NEXT:    v_readfirstlane_b32 s57, v13
; VI-NEXT:    v_readfirstlane_b32 s58, v10
; VI-NEXT:    v_readfirstlane_b32 s41, v9
; VI-NEXT:    v_readfirstlane_b32 s46, v8
; VI-NEXT:    v_readfirstlane_b32 s13, v7
; VI-NEXT:    v_readfirstlane_b32 s42, v6
; VI-NEXT:    v_readfirstlane_b32 s10, v5
; VI-NEXT:    v_readfirstlane_b32 s14, v4
; VI-NEXT:    v_readfirstlane_b32 s7, v3
; VI-NEXT:    v_readfirstlane_b32 s11, v2
; VI-NEXT:    v_readfirstlane_b32 s6, v1
; VI-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_ushort v13, off, s[0:3], s32
; VI-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:328
; VI-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:324
; VI-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:320
; VI-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:316
; VI-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:312
; VI-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:308
; VI-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:304
; VI-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:300
; VI-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:296
; VI-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:292
; VI-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:288
; VI-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:284
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:272
; VI-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:268
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:264
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:260
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:256
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:252
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:248
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:244
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:236
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    s_and_b64 s[4:5], vcc, exec
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:216
; VI-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:212
; VI-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:208
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:200
; VI-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:196
; VI-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:192
; VI-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:188
; VI-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:184
; VI-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:180
; VI-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:176
; VI-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:172
; VI-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:168
; VI-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:164
; VI-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:160
; VI-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:156
; VI-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:152
; VI-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:148
; VI-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:144
; VI-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:140
; VI-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:136
; VI-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:132
; VI-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:128
; VI-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:124
; VI-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:120
; VI-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:116
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:112
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:108
; VI-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:104
; VI-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:100
; VI-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:96
; VI-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:92
; VI-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:88
; VI-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:84
; VI-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:80
; VI-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:76
; VI-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:72
; VI-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:68
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:64
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:60
; VI-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:56
; VI-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:52
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:48
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(5)
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:44
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:40
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:36
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:28
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:24
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:20
; VI-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    s_cbranch_scc0 .LBB97_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_mov_b32_e32 v11, 0xc0c0004
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, s19
; VI-NEXT:    v_mov_b32_e32 v3, s23
; VI-NEXT:    v_mov_b32_e32 v0, s17
; VI-NEXT:    v_perm_b32 v1, s18, v1, v11
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v2, s21
; VI-NEXT:    v_perm_b32 v3, s22, v3, v11
; VI-NEXT:    v_perm_b32 v0, s16, v0, v11
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_perm_b32 v2, s20, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v0, v0, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s27
; VI-NEXT:    v_mov_b32_e32 v2, s25
; VI-NEXT:    v_perm_b32 v3, s26, v3, v11
; VI-NEXT:    v_perm_b32 v2, s24, v2, v11
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_mov_b32_e32 v4, s6
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_mov_b32_e32 v3, s29
; VI-NEXT:    v_perm_b32 v4, s8, v4, v11
; VI-NEXT:    v_perm_b32 v3, s28, v3, v11
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_mov_b32_e32 v5, s10
; VI-NEXT:    v_or_b32_e32 v3, v3, v4
; VI-NEXT:    v_mov_b32_e32 v4, s7
; VI-NEXT:    v_perm_b32 v5, s14, v5, v11
; VI-NEXT:    v_perm_b32 v4, s11, v4, v11
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_mov_b32_e32 v6, s41
; VI-NEXT:    v_or_b32_e32 v4, v4, v5
; VI-NEXT:    v_mov_b32_e32 v5, s13
; VI-NEXT:    v_perm_b32 v6, s46, v6, v11
; VI-NEXT:    v_perm_b32 v5, s42, v5, v11
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_mov_b32_e32 v7, s57
; VI-NEXT:    v_or_b32_e32 v5, v5, v6
; VI-NEXT:    v_mov_b32_e32 v6, s45
; VI-NEXT:    v_perm_b32 v7, s63, v7, v11
; VI-NEXT:    v_perm_b32 v6, s58, v6, v11
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_mov_b32_e32 v8, s61
; VI-NEXT:    v_or_b32_e32 v6, v6, v7
; VI-NEXT:    v_mov_b32_e32 v7, s62
; VI-NEXT:    v_perm_b32 v8, s73, v8, v11
; VI-NEXT:    v_perm_b32 v7, s74, v7, v11
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_mov_b32_e32 v9, s12
; VI-NEXT:    v_or_b32_e32 v7, v7, v8
; VI-NEXT:    v_mov_b32_e32 v8, s9
; VI-NEXT:    v_perm_b32 v9, s40, v9, v11
; VI-NEXT:    v_perm_b32 v8, s76, v8, v11
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_mov_b32_e32 v43, v10
; VI-NEXT:    v_mov_b32_e32 v10, s43
; VI-NEXT:    v_or_b32_e32 v8, v8, v9
; VI-NEXT:    v_mov_b32_e32 v9, s15
; VI-NEXT:    v_perm_b32 v10, s56, v10, v11
; VI-NEXT:    v_perm_b32 v9, s44, v9, v11
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_mov_b32_e32 v35, v12
; VI-NEXT:    v_mov_b32_e32 v12, s59
; VI-NEXT:    v_or_b32_e32 v9, v9, v10
; VI-NEXT:    v_mov_b32_e32 v10, s47
; VI-NEXT:    v_perm_b32 v12, s72, v12, v11
; VI-NEXT:    v_perm_b32 v10, s60, v10, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    v_or_b32_e32 v10, v10, v12
; VI-NEXT:    v_perm_b32 v12, v15, v14, s4
; VI-NEXT:    v_perm_b32 v11, s75, v13, v11
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    v_or_b32_e32 v11, v11, v12
; VI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v18, v23, v18, s4
; VI-NEXT:    v_perm_b32 v23, v51, v33, s4
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_perm_b32 v12, v56, v12, s4
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v13, v47, v13, s4
; VI-NEXT:    v_or_b32_e32 v12, v13, v12
; VI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v14, v59, v14, s4
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v15, v32, v15, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v61, v13, s4
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; VI-NEXT:    v_or_b32_e32 v13, v14, v13
; VI-NEXT:    v_perm_b32 v14, v17, v60, s4
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; VI-NEXT:    v_or_b32_e32 v14, v15, v14
; VI-NEXT:    v_perm_b32 v15, v21, v16, s4
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v17, v28, v24, s4
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_perm_b32 v21, v58, v46, s4
; VI-NEXT:    v_perm_b32 v24, v36, v49, s4
; VI-NEXT:    v_mov_b32_e32 v28, v52
; VI-NEXT:    v_mov_b32_e32 v52, v53
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v16, v19, v16, s4
; VI-NEXT:    v_perm_b32 v19, v26, v22, s4
; VI-NEXT:    v_or_b32_e32 v17, v19, v17
; VI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_e32 v15, v16, v15
; VI-NEXT:    v_perm_b32 v16, v25, v20, s4
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; VI-NEXT:    v_or_b32_e32 v16, v18, v16
; VI-NEXT:    v_perm_b32 v18, v45, v27, s4
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; VI-NEXT:    v_perm_b32 v20, v44, v29, s4
; VI-NEXT:    v_perm_b32 v22, v53, v42, s4
; VI-NEXT:    v_mov_b32_e32 v45, v34
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(3)
; VI-NEXT:    v_perm_b32 v19, v30, v19, s4
; VI-NEXT:    v_or_b32_e32 v18, v19, v18
; VI-NEXT:    v_perm_b32 v19, v55, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_or_b32_e32 v19, v20, v19
; VI-NEXT:    v_perm_b32 v20, v39, v57, s4
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_e32 v20, v21, v20
; VI-NEXT:    v_perm_b32 v21, v38, v63, s4
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_or_b32_e32 v21, v22, v21
; VI-NEXT:    v_perm_b32 v22, v54, v37, s4
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_or_b32_e32 v22, v23, v22
; VI-NEXT:    v_perm_b32 v23, v40, v50, s4
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_or_b32_e32 v23, v24, v23
; VI-NEXT:    v_perm_b32 v24, v48, v34, s4
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_mov_b32_e32 v48, v25
; VI-NEXT:    v_perm_b32 v25, v25, v41, s4
; VI-NEXT:    v_or_b32_e32 v24, v25, v24
; VI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v34, v33
; VI-NEXT:    v_mov_b32_e32 v33, v26
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v55, v57
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_mov_b32_e32 v53, v25
; VI-NEXT:    v_perm_b32 v25, v26, v25, s4
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    v_or_b32_e32 v25, v26, v25
; VI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v27, v26, s4
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v29, v27, s4
; VI-NEXT:    v_or_b32_e32 v26, v27, v26
; VI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v29, v27, s4
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v28, v29, s4
; VI-NEXT:    v_or_b32_e32 v27, v28, v27
; VI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v28, v28, v35, s4
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v62, v29, s4
; VI-NEXT:    v_or_b32_e32 v28, v29, v28
; VI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, v44
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v43, v29, s4
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; VI-NEXT:    v_or_b32_e32 v29, v30, v29
; VI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v30, v31, v30, s4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v35, v31, s4
; VI-NEXT:    v_or_b32_e32 v30, v31, v30
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v31, v35, v31, s4
; VI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; VI-NEXT:    v_mov_b32_e32 v35, v43
; VI-NEXT:    v_perm_b32 v43, v44, v43, s4
; VI-NEXT:    v_or_b32_e32 v31, v43, v31
; VI-NEXT:    v_mov_b32_e32 v43, v39
; VI-NEXT:    v_mov_b32_e32 v39, v34
; VI-NEXT:    v_mov_b32_e32 v34, v52
; VI-NEXT:    v_mov_b32_e32 v52, v45
; VI-NEXT:    s_cbranch_execnz .LBB97_3
; VI-NEXT:  .LBB97_2: ; %cmp.true
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_add_i32 s75, s75, 3
; VI-NEXT:    s_add_i32 s72, s72, 3
; VI-NEXT:    s_add_i32 s60, s60, 3
; VI-NEXT:    s_add_i32 s56, s56, 3
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_add_i32 s76, s76, 3
; VI-NEXT:    s_add_i32 s73, s73, 3
; VI-NEXT:    s_add_i32 s74, s74, 3
; VI-NEXT:    s_add_i32 s63, s63, 3
; VI-NEXT:    s_add_i32 s58, s58, 3
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    v_mov_b32_e32 v18, s17
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v62
; VI-NEXT:    v_perm_b32 v0, v0, v35, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, s21
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v35, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
; VI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v27, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x300, v27
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v26, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v28, v0, v1, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v33
; VI-NEXT:    v_perm_b32 v25, v0, v53, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x300, v28
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; VI-NEXT:    v_or_b32_sdwa v26, v26, v28 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v26, vcc, 0x3000000, v26
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v29, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x300, v29
; VI-NEXT:    v_or_b32_sdwa v25, v25, v29 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x3000000, v25
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_perm_b32 v24, v0, v52, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v48
; VI-NEXT:    v_perm_b32 v30, v0, v41, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v40
; VI-NEXT:    v_perm_b32 v23, v0, v50, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v36
; VI-NEXT:    v_perm_b32 v31, v0, v49, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v54
; VI-NEXT:    v_perm_b32 v22, v0, v37, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v51
; VI-NEXT:    v_perm_b32 v33, v0, v39, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v38
; VI-NEXT:    v_perm_b32 v21, v0, v63, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v34
; VI-NEXT:    v_perm_b32 v36, v0, v42, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v43
; VI-NEXT:    v_perm_b32 v20, v0, v55, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v58
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v19, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v63, s19
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x300, v30
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x300, v31
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x300, v33
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; VI-NEXT:    v_or_b32_sdwa v22, v22, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v23, v23, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v24, v24, v30 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v22, vcc, 0x3000000, v22
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x3000000, v23
; VI-NEXT:    v_add_u32_e32 v24, vcc, 0x3000000, v24
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v52, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v53, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v37, vcc, 0x300, v53
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v38, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v34, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v10, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v11, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v12, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v13, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v14, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v15, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x300, v15
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    v_perm_b32 v16, v0, v60, s4
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v32
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v17, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v61
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x300, v34
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v54, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v59
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v40, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v56
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v42, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v47
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v43, v0, v1, s4
; VI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_add_u32_e32 v0, vcc, 3, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v44, v0, v1, s4
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v0, 0xc0c0004
; VI-NEXT:    v_perm_b32 v62, s20, v62, v0
; VI-NEXT:    v_perm_b32 v63, s18, v63, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v45, s75, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s59
; VI-NEXT:    v_perm_b32 v55, s72, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s47
; VI-NEXT:    v_perm_b32 v41, s60, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s43
; VI-NEXT:    v_perm_b32 v9, s56, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s15
; VI-NEXT:    v_perm_b32 v46, s44, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s12
; VI-NEXT:    v_perm_b32 v8, s40, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s9
; VI-NEXT:    v_perm_b32 v47, s76, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s61
; VI-NEXT:    v_perm_b32 v7, s73, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s62
; VI-NEXT:    v_perm_b32 v56, s74, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s57
; VI-NEXT:    v_perm_b32 v6, s63, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s45
; VI-NEXT:    v_perm_b32 v57, s58, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s41
; VI-NEXT:    v_perm_b32 v5, s46, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s13
; VI-NEXT:    v_perm_b32 v58, s42, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s10
; VI-NEXT:    v_perm_b32 v4, s14, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s7
; VI-NEXT:    v_perm_b32 v59, s11, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s6
; VI-NEXT:    v_perm_b32 v3, s8, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s29
; VI-NEXT:    v_perm_b32 v60, s28, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s27
; VI-NEXT:    v_perm_b32 v2, s26, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s25
; VI-NEXT:    v_perm_b32 v61, s24, v1, v0
; VI-NEXT:    v_mov_b32_e32 v1, s23
; VI-NEXT:    v_perm_b32 v1, s22, v1, v0
; VI-NEXT:    v_perm_b32 v0, s16, v18, v0
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v63
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x300, v0
; VI-NEXT:    v_or_b32_sdwa v0, v18, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v62
; VI-NEXT:    v_or_b32_sdwa v1, v1, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v61
; VI-NEXT:    v_or_b32_sdwa v2, v2, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v60
; VI-NEXT:    v_or_b32_sdwa v3, v3, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v59
; VI-NEXT:    v_or_b32_sdwa v4, v4, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v58
; VI-NEXT:    v_or_b32_sdwa v5, v5, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v57
; VI-NEXT:    v_or_b32_sdwa v6, v6, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v56
; VI-NEXT:    v_or_b32_sdwa v7, v7, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v47
; VI-NEXT:    v_or_b32_sdwa v8, v8, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x300, v46
; VI-NEXT:    v_or_b32_sdwa v9, v9, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v10
; VI-NEXT:    v_add_u32_e32 v46, vcc, 0x300, v11
; VI-NEXT:    v_add_u32_e32 v56, vcc, 0x300, v13
; VI-NEXT:    v_lshlrev_b32_e32 v57, 16, v14
; VI-NEXT:    v_lshlrev_b32_e32 v14, 16, v16
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x300, v17
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v54
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x300, v40
; VI-NEXT:    v_or_b32_sdwa v13, v13, v17 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v17, v18, v46 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v47, 16, v12
; VI-NEXT:    v_lshlrev_b32_e32 v12, 16, v42
; VI-NEXT:    v_add_u32_e32 v54, vcc, 0x300, v43
; VI-NEXT:    v_or_b32_sdwa v12, v12, v54 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v55
; VI-NEXT:    v_add_u32_e32 v55, vcc, 0x300, v41
; VI-NEXT:    v_or_b32_sdwa v10, v10, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v44
; VI-NEXT:    v_add_u32_e32 v40, vcc, 0x300, v45
; VI-NEXT:    v_or_b32_sdwa v11, v11, v40 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v43, vcc, 0x300, v35
; VI-NEXT:    v_add_u32_e32 v35, vcc, 0x300, v36
; VI-NEXT:    v_add_u32_e32 v36, vcc, 0x300, v19
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v52
; VI-NEXT:    v_or_b32_sdwa v14, v14, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v15, v57, v15 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v16, v47, v56 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v19, v19, v37 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v20, v20, v36 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v21, v21, v35 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x3000000, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x3000000, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x3000000, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x3000000, v3
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x3000000, v4
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x3000000, v5
; VI-NEXT:    v_add_u32_e32 v6, vcc, 0x3000000, v6
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x3000000, v7
; VI-NEXT:    v_add_u32_e32 v8, vcc, 0x3000000, v8
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x3000000, v9
; VI-NEXT:    v_add_u32_e32 v10, vcc, 0x3000000, v10
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x3000000, v11
; VI-NEXT:    v_add_u32_e32 v12, vcc, 0x3000000, v12
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x3000000, v13
; VI-NEXT:    v_add_u32_e32 v14, vcc, 0x3000000, v14
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x3000000, v15
; VI-NEXT:    v_add_u32_e32 v16, vcc, 0x3000000, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x3000000, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x3000000, v19
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x3000000, v20
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x3000000, v21
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v54, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v55, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v31, v54, v55 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x3000000, v31
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v40, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v41, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v30, v40, v41 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v30, vcc, 0x3000000, v30
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v42, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v29, v42, v43 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x3000000, v29
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v44, 16, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u32_e32 v45, vcc, 0x300, v18
; VI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    v_or_b32_sdwa v28, v44, v45 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v28, vcc, 0x3000000, v28
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_lshlrev_b32_e32 v46, 16, v18
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v38
; VI-NEXT:    v_or_b32_sdwa v18, v18, v32 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_or_b32_sdwa v27, v46, v27 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x3000000, v18
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x3000000, v27
; VI-NEXT:  .LBB97_3: ; %end
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB97_4:
; VI-NEXT:    v_mov_b32_e32 v43, v39
; VI-NEXT:    v_mov_b32_e32 v52, v34
; VI-NEXT:    v_mov_b32_e32 v39, v33
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v34, v53
; VI-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; VI-NEXT:    v_mov_b32_e32 v62, v2
; VI-NEXT:    v_mov_b32_e32 v35, v1
; VI-NEXT:    v_mov_b32_e32 v55, v57
; VI-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; VI-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; VI-NEXT:    s_branch .LBB97_2
;
; GFX9-LABEL: bitcast_v128i8_to_v64i16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s61, v11
; GFX9-NEXT:    v_readfirstlane_b32 s63, v10
; GFX9-NEXT:    v_readfirstlane_b32 s7, v0
; GFX9-NEXT:    buffer_load_ushort v10, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_ushort v11, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_ushort v12, off, s[0:3], s32
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:328
; GFX9-NEXT:    v_readfirstlane_b32 s43, v2
; GFX9-NEXT:    v_readfirstlane_b32 s6, v1
; GFX9-NEXT:    v_readfirstlane_b32 s8, v5
; GFX9-NEXT:    v_readfirstlane_b32 s10, v4
; GFX9-NEXT:    v_readfirstlane_b32 s15, v3
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_readfirstlane_b32 s12, v9
; GFX9-NEXT:    v_readfirstlane_b32 s14, v8
; GFX9-NEXT:    v_readfirstlane_b32 s47, v7
; GFX9-NEXT:    v_readfirstlane_b32 s59, v6
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:332
; GFX9-NEXT:    v_readfirstlane_b32 s42, v13
; GFX9-NEXT:    v_readfirstlane_b32 s74, v30
; GFX9-NEXT:    v_readfirstlane_b32 s44, v29
; GFX9-NEXT:    v_readfirstlane_b32 s56, v28
; GFX9-NEXT:    v_readfirstlane_b32 s62, v27
; GFX9-NEXT:    v_readfirstlane_b32 s72, v26
; GFX9-NEXT:    v_readfirstlane_b32 s13, v25
; GFX9-NEXT:    v_readfirstlane_b32 s40, v24
; GFX9-NEXT:    v_readfirstlane_b32 s57, v23
; GFX9-NEXT:    v_readfirstlane_b32 s60, v22
; GFX9-NEXT:    v_readfirstlane_b32 s9, v21
; GFX9-NEXT:    v_readfirstlane_b32 s11, v20
; GFX9-NEXT:    v_readfirstlane_b32 s41, v19
; GFX9-NEXT:    v_readfirstlane_b32 s76, v18
; GFX9-NEXT:    v_readfirstlane_b32 s45, v17
; GFX9-NEXT:    v_readfirstlane_b32 s58, v16
; GFX9-NEXT:    v_readfirstlane_b32 s73, v15
; GFX9-NEXT:    v_readfirstlane_b32 s75, v14
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:324
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:320
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:316
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v2, off, s[0:3], s32 offset:312
; GFX9-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:308
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:304
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:300
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v4, off, s[0:3], s32 offset:296
; GFX9-NEXT:    buffer_load_ushort v5, off, s[0:3], s32 offset:292
; GFX9-NEXT:    buffer_load_ushort v3, off, s[0:3], s32 offset:288
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:284
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v8, off, s[0:3], s32 offset:280
; GFX9-NEXT:    buffer_load_ushort v7, off, s[0:3], s32 offset:276
; GFX9-NEXT:    buffer_load_ushort v6, off, s[0:3], s32 offset:272
; GFX9-NEXT:    buffer_load_ushort v63, off, s[0:3], s32 offset:268
; GFX9-NEXT:    buffer_load_ushort v48, off, s[0:3], s32 offset:264
; GFX9-NEXT:    buffer_load_ushort v54, off, s[0:3], s32 offset:260
; GFX9-NEXT:    buffer_load_ushort v61, off, s[0:3], s32 offset:256
; GFX9-NEXT:    buffer_load_ushort v33, off, s[0:3], s32 offset:252
; GFX9-NEXT:    buffer_load_ushort v9, off, s[0:3], s32 offset:248
; GFX9-NEXT:    buffer_load_ushort v62, off, s[0:3], s32 offset:244
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:240
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v43, off, s[0:3], s32 offset:236
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:232
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:228
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:224
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:220
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:216
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:212
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:208
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:204
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:200
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:196
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    s_and_b64 s[4:5], vcc, exec
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 offset:192
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:188
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_ushort v47, off, s[0:3], s32 offset:184
; GFX9-NEXT:    buffer_load_ushort v59, off, s[0:3], s32 offset:180
; GFX9-NEXT:    buffer_load_ushort v42, off, s[0:3], s32 offset:176
; GFX9-NEXT:    buffer_load_ushort v58, off, s[0:3], s32 offset:172
; GFX9-NEXT:    buffer_load_ushort v44, off, s[0:3], s32 offset:168
; GFX9-NEXT:    buffer_load_ushort v56, off, s[0:3], s32 offset:164
; GFX9-NEXT:    buffer_load_ushort v60, off, s[0:3], s32 offset:160
; GFX9-NEXT:    buffer_load_ushort v53, off, s[0:3], s32 offset:156
; GFX9-NEXT:    buffer_load_ushort v49, off, s[0:3], s32 offset:152
; GFX9-NEXT:    buffer_load_ushort v45, off, s[0:3], s32 offset:148
; GFX9-NEXT:    buffer_load_ushort v57, off, s[0:3], s32 offset:144
; GFX9-NEXT:    buffer_load_ushort v55, off, s[0:3], s32 offset:140
; GFX9-NEXT:    buffer_load_ushort v38, off, s[0:3], s32 offset:136
; GFX9-NEXT:    buffer_load_ushort v32, off, s[0:3], s32 offset:132
; GFX9-NEXT:    buffer_load_ushort v46, off, s[0:3], s32 offset:128
; GFX9-NEXT:    buffer_load_ushort v40, off, s[0:3], s32 offset:124
; GFX9-NEXT:    buffer_load_ushort v31, off, s[0:3], s32 offset:120
; GFX9-NEXT:    buffer_load_ushort v52, off, s[0:3], s32 offset:116
; GFX9-NEXT:    buffer_load_ushort v50, off, s[0:3], s32 offset:112
; GFX9-NEXT:    buffer_load_ushort v41, off, s[0:3], s32 offset:108
; GFX9-NEXT:    buffer_load_ushort v27, off, s[0:3], s32 offset:104
; GFX9-NEXT:    buffer_load_ushort v37, off, s[0:3], s32 offset:100
; GFX9-NEXT:    buffer_load_ushort v36, off, s[0:3], s32 offset:96
; GFX9-NEXT:    buffer_load_ushort v51, off, s[0:3], s32 offset:92
; GFX9-NEXT:    buffer_load_ushort v23, off, s[0:3], s32 offset:88
; GFX9-NEXT:    buffer_load_ushort v28, off, s[0:3], s32 offset:84
; GFX9-NEXT:    buffer_load_ushort v34, off, s[0:3], s32 offset:80
; GFX9-NEXT:    buffer_load_ushort v39, off, s[0:3], s32 offset:76
; GFX9-NEXT:    buffer_load_ushort v19, off, s[0:3], s32 offset:72
; GFX9-NEXT:    buffer_load_ushort v24, off, s[0:3], s32 offset:68
; GFX9-NEXT:    buffer_load_ushort v29, off, s[0:3], s32 offset:64
; GFX9-NEXT:    buffer_load_ushort v35, off, s[0:3], s32 offset:60
; GFX9-NEXT:    buffer_load_ushort v16, off, s[0:3], s32 offset:56
; GFX9-NEXT:    buffer_load_ushort v20, off, s[0:3], s32 offset:52
; GFX9-NEXT:    buffer_load_ushort v25, off, s[0:3], s32 offset:48
; GFX9-NEXT:    buffer_load_ushort v30, off, s[0:3], s32 offset:44
; GFX9-NEXT:    buffer_load_ushort v14, off, s[0:3], s32 offset:40
; GFX9-NEXT:    buffer_load_ushort v17, off, s[0:3], s32 offset:36
; GFX9-NEXT:    buffer_load_ushort v21, off, s[0:3], s32 offset:32
; GFX9-NEXT:    buffer_load_ushort v26, off, s[0:3], s32 offset:28
; GFX9-NEXT:    buffer_load_ushort v13, off, s[0:3], s32 offset:24
; GFX9-NEXT:    buffer_load_ushort v15, off, s[0:3], s32 offset:20
; GFX9-NEXT:    buffer_load_ushort v18, off, s[0:3], s32 offset:16
; GFX9-NEXT:    buffer_load_ushort v22, off, s[0:3], s32 offset:12
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(6)
; GFX9-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(12)
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; GFX9-NEXT:    s_cbranch_scc0 .LBB97_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_mov_b32_e32 v0, s17
; GFX9-NEXT:    v_mov_b32_e32 v8, 0xc0c0004
; GFX9-NEXT:    v_mov_b32_e32 v2, s21
; GFX9-NEXT:    v_perm_b32 v0, s16, v0, v8
; GFX9-NEXT:    v_mov_b32_e32 v1, s19
; GFX9-NEXT:    v_perm_b32 v2, s20, v2, v8
; GFX9-NEXT:    v_mov_b32_e32 v3, s23
; GFX9-NEXT:    v_perm_b32 v1, s18, v1, v8
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX9-NEXT:    v_perm_b32 v3, s22, v3, v8
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX9-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX9-NEXT:    v_mov_b32_e32 v2, s25
; GFX9-NEXT:    v_perm_b32 v2, s24, v2, v8
; GFX9-NEXT:    v_mov_b32_e32 v3, s27
; GFX9-NEXT:    v_perm_b32 v3, s26, v3, v8
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
; GFX9-NEXT:    v_mov_b32_e32 v3, s29
; GFX9-NEXT:    v_perm_b32 v3, s28, v3, v8
; GFX9-NEXT:    v_mov_b32_e32 v4, s6
; GFX9-NEXT:    v_perm_b32 v4, s7, v4, v8
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
; GFX9-NEXT:    v_mov_b32_e32 v4, s15
; GFX9-NEXT:    v_perm_b32 v4, s43, v4, v8
; GFX9-NEXT:    v_mov_b32_e32 v5, s8
; GFX9-NEXT:    v_perm_b32 v5, s10, v5, v8
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX9-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX9-NEXT:    v_mov_b32_e32 v5, s47
; GFX9-NEXT:    v_perm_b32 v5, s59, v5, v8
; GFX9-NEXT:    v_mov_b32_e32 v6, s12
; GFX9-NEXT:    v_perm_b32 v6, s14, v6, v8
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX9-NEXT:    v_mov_b32_e32 v6, s61
; GFX9-NEXT:    v_perm_b32 v6, s63, v6, v8
; GFX9-NEXT:    v_mov_b32_e32 v7, s42
; GFX9-NEXT:    v_perm_b32 v7, s46, v7, v8
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    v_lshl_or_b32 v6, v7, 16, v6
; GFX9-NEXT:    v_mov_b32_e32 v7, s73
; GFX9-NEXT:    v_perm_b32 v7, s75, v7, v8
; GFX9-NEXT:    v_mov_b32_e32 v63, v33
; GFX9-NEXT:    v_mov_b32_e32 v33, v9
; GFX9-NEXT:    v_mov_b32_e32 v9, s45
; GFX9-NEXT:    v_perm_b32 v9, s58, v9, v8
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX9-NEXT:    v_lshl_or_b32 v7, v9, 16, v7
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v10
; GFX9-NEXT:    v_perm_b32 v10, s74, v12, v8
; GFX9-NEXT:    v_or_b32_sdwa v9, v11, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v11, v9, 16, v10
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v18
; GFX9-NEXT:    v_or_b32_sdwa v9, v22, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v13
; GFX9-NEXT:    v_or_b32_sdwa v10, v15, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_lshl_or_b32 v12, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v14
; GFX9-NEXT:    v_perm_b32 v10, v26, v21, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v17, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v13, v9, 16, v10
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v25
; GFX9-NEXT:    v_or_b32_sdwa v9, v30, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v16
; GFX9-NEXT:    v_or_b32_sdwa v10, v20, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v14, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v19
; GFX9-NEXT:    v_perm_b32 v10, v35, v29, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v24, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v15, v9, 16, v10
; GFX9-NEXT:    v_perm_b32 v9, v39, v34, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v51, v36, s4
; GFX9-NEXT:    v_perm_b32 v16, v28, v23, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_perm_b32 v17, v37, v27, s4
; GFX9-NEXT:    v_lshl_or_b32 v16, v16, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v41, v50, s4
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v10
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v52, v31, s4
; GFX9-NEXT:    v_lshl_or_b32 v18, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v40, v46, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v32, v38, s4
; GFX9-NEXT:    v_lshl_or_b32 v19, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v55, v57, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v45, v49, s4
; GFX9-NEXT:    v_lshl_or_b32 v20, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v53, v60, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v56, v44, s4
; GFX9-NEXT:    v_lshl_or_b32 v21, v10, 16, v9
; GFX9-NEXT:    v_perm_b32 v9, v58, v42, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_perm_b32 v10, v59, v47, s4
; GFX9-NEXT:    v_lshl_or_b32 v22, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v50, v32
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v51, v49
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v41, v46
; GFX9-NEXT:    v_mov_b32_e32 v39, v52
; GFX9-NEXT:    v_mov_b32_e32 v52, v38
; GFX9-NEXT:    v_mov_b32_e32 v34, v37
; GFX9-NEXT:    v_mov_b32_e32 v37, v31
; GFX9-NEXT:    s_lshl_b32 s5, s13, 8
; GFX9-NEXT:    s_lshl_b32 s77, s9, 8
; GFX9-NEXT:    s_lshl_b32 s78, s44, 8
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(9)
; GFX9-NEXT:    v_perm_b32 v9, v9, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(7)
; GFX9-NEXT:    v_perm_b32 v10, v32, v58, s4
; GFX9-NEXT:    v_lshl_or_b32 v23, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v49
; GFX9-NEXT:    v_or_b32_sdwa v10, v24, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v24, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v35
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v25, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v25, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v33
; GFX9-NEXT:    v_or_b32_sdwa v10, v62, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v33, v46
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    v_or_b32_sdwa v9, v43, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_lshl_or_b32 v26, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v48
; GFX9-NEXT:    v_perm_b32 v10, v63, v61, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v54, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v27, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v54, v61
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_mov_b32_e32 v48, v63
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
; GFX9-NEXT:    v_or_b32_sdwa v10, v28, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshl_or_b32 v28, v10, 16, v9
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_or_b32_sdwa v9, v10, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v36, v10, s4
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v29, v9, 16, v10
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v43
; GFX9-NEXT:    v_or_b32_sdwa v9, v62, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 8, v10
; GFX9-NEXT:    v_or_b32_sdwa v10, v30, v10 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_lshl_or_b32 v30, v10, 16, v9
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 8, v38
; GFX9-NEXT:    v_perm_b32 v10, v63, v61, s4
; GFX9-NEXT:    v_or_b32_sdwa v9, v46, v9 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v31, v9, 16, v10
; GFX9-NEXT:    v_mov_b32_e32 v9, s57
; GFX9-NEXT:    v_perm_b32 v8, s60, v9, v8
; GFX9-NEXT:    s_and_b32 s4, s40, 0xff
; GFX9-NEXT:    s_or_b32 s4, s4, s5
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX9-NEXT:    v_lshl_or_b32 v9, s4, 16, v8
; GFX9-NEXT:    s_and_b32 s4, s76, 0xff
; GFX9-NEXT:    s_lshl_b32 s5, s41, 8
; GFX9-NEXT:    s_or_b32 s4, s4, s5
; GFX9-NEXT:    s_and_b32 s5, s11, 0xff
; GFX9-NEXT:    s_or_b32 s5, s5, s77
; GFX9-NEXT:    s_pack_ll_b32_b16 s4, s4, s5
; GFX9-NEXT:    s_and_b32 s5, s72, 0xff
; GFX9-NEXT:    s_lshl_b32 s77, s62, 8
; GFX9-NEXT:    s_or_b32 s5, s5, s77
; GFX9-NEXT:    s_and_b32 s77, s56, 0xff
; GFX9-NEXT:    s_or_b32 s77, s77, s78
; GFX9-NEXT:    s_pack_ll_b32_b16 s5, s5, s77
; GFX9-NEXT:    v_mov_b32_e32 v8, s4
; GFX9-NEXT:    v_mov_b32_e32 v10, s5
; GFX9-NEXT:    s_mov_b64 s[4:5], 0
; GFX9-NEXT:    v_mov_b32_e32 v63, v36
; GFX9-NEXT:    s_branch .LBB97_3
; GFX9-NEXT:  .LBB97_2:
; GFX9-NEXT:    v_mov_b32_e32 v54, v61
; GFX9-NEXT:    v_mov_b32_e32 v50, v32
; GFX9-NEXT:    v_mov_b32_e32 v51, v49
; GFX9-NEXT:    v_mov_b32_e32 v39, v52
; GFX9-NEXT:    v_mov_b32_e32 v52, v38
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v41, v46
; GFX9-NEXT:    v_mov_b32_e32 v34, v37
; GFX9-NEXT:    v_mov_b32_e32 v37, v31
; GFX9-NEXT:    s_mov_b64 s[4:5], -1
; GFX9-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX9-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX9-NEXT:  .LBB97_3: ; %Flow
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; GFX9-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; GFX9-NEXT:    s_cbranch_vccnz .LBB97_5
; GFX9-NEXT:  ; %bb.4: ; %cmp.true
; GFX9-NEXT:    s_waitcnt vmcnt(5)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v48
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v0, v0, v61, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v33
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v0, v0, v38, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v62
; GFX9-NEXT:    v_perm_b32 v0, v0, v43, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    s_add_i32 s74, s74, 3
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    s_add_i32 s72, s72, 3
; GFX9-NEXT:    s_add_i32 s56, s56, 3
; GFX9-NEXT:    s_add_i32 s60, s60, 3
; GFX9-NEXT:    s_add_i32 s40, s40, 3
; GFX9-NEXT:    s_add_i32 s76, s76, 3
; GFX9-NEXT:    s_add_i32 s11, s11, 3
; GFX9-NEXT:    s_add_i32 s75, s75, 3
; GFX9-NEXT:    s_add_i32 s58, s58, 3
; GFX9-NEXT:    s_add_i32 s63, s63, 3
; GFX9-NEXT:    s_add_i32 s46, s46, 3
; GFX9-NEXT:    s_add_i32 s59, s59, 3
; GFX9-NEXT:    s_add_i32 s14, s14, 3
; GFX9-NEXT:    s_add_i32 s43, s43, 3
; GFX9-NEXT:    s_add_i32 s10, s10, 3
; GFX9-NEXT:    s_add_i32 s28, s28, 3
; GFX9-NEXT:    s_add_i32 s7, s7, 3
; GFX9-NEXT:    s_add_i32 s24, s24, 3
; GFX9-NEXT:    s_add_i32 s26, s26, 3
; GFX9-NEXT:    s_add_i32 s20, s20, 3
; GFX9-NEXT:    s_add_i32 s22, s22, 3
; GFX9-NEXT:    s_add_i32 s16, s16, 3
; GFX9-NEXT:    v_mov_b32_e32 v33, s17
; GFX9-NEXT:    s_add_i32 s18, s18, 3
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v63
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v38, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v62, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v54, s4
; GFX9-NEXT:    v_add_u32_e32 v63, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v31, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v30, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v35, s4
; GFX9-NEXT:    v_add_u32_e32 v29, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v36
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v28, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v49, s4
; GFX9-NEXT:    v_add_u32_e32 v27, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v26, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v32
; GFX9-NEXT:    v_perm_b32 v0, v0, v58, s4
; GFX9-NEXT:    v_add_u32_e32 v25, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; GFX9-NEXT:    v_mov_b32_e32 v32, s23
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v42, s4
; GFX9-NEXT:    v_add_u32_e32 v24, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v59
; GFX9-NEXT:    v_perm_b32 v0, v0, v47, s4
; GFX9-NEXT:    v_add_u32_e32 v23, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v53
; GFX9-NEXT:    v_perm_b32 v0, v0, v60, s4
; GFX9-NEXT:    v_add_u32_e32 v22, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v56
; GFX9-NEXT:    v_perm_b32 v0, v0, v44, s4
; GFX9-NEXT:    v_add_u32_e32 v21, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v55
; GFX9-NEXT:    v_perm_b32 v0, v0, v57, s4
; GFX9-NEXT:    v_add_u32_e32 v20, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v45
; GFX9-NEXT:    v_perm_b32 v0, v0, v51, s4
; GFX9-NEXT:    v_add_u32_e32 v49, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v40
; GFX9-NEXT:    v_perm_b32 v0, v0, v41, s4
; GFX9-NEXT:    v_add_u32_e32 v19, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v50
; GFX9-NEXT:    v_perm_b32 v0, v0, v52, s4
; GFX9-NEXT:    v_add_u32_e32 v52, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX9-NEXT:    v_lshl_or_b32 v21, v21, 16, v22
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff, v24
; GFX9-NEXT:    v_lshl_or_b32 v22, v23, 16, v22
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff, v26
; GFX9-NEXT:    v_lshl_or_b32 v23, v25, 16, v23
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff, v30
; GFX9-NEXT:    v_lshl_or_b32 v25, v29, 16, v25
; GFX9-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff, v28
; GFX9-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v24, v27, 16, v24
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff, v63
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; GFX9-NEXT:    v_lshl_or_b32 v19, v52, 16, v19
; GFX9-NEXT:    v_lshl_or_b32 v20, v49, 16, v20
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v18, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v39
; GFX9-NEXT:    v_perm_b32 v0, v0, v37, s4
; GFX9-NEXT:    v_add_u32_e32 v37, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; GFX9-NEXT:    v_lshl_or_b32 v18, v37, 16, v18
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v54, 0x300, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v34
; GFX9-NEXT:    v_mov_b32_e32 v34, s19
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v17, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v16, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v55, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v16, v55, 16, v16
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v15, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v40, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v15, v40, 16, v15
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v14, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v41, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v14, v41, 16, v14
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v13, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v42, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v13, v42, 16, v13
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    v_add_u32_e32 v12, 0x300, v0
; GFX9-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_add_u32_e32 v0, 3, v0
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v0, v0, v1, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; GFX9-NEXT:    v_add_u32_e32 v43, 0x300, v0
; GFX9-NEXT:    v_mov_b32_e32 v0, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v32, s22, v32, v0
; GFX9-NEXT:    v_add_u32_e32 v32, 0x300, v32
; GFX9-NEXT:    v_perm_b32 v33, s16, v33, v0
; GFX9-NEXT:    v_add_u32_e32 v33, 0x300, v33
; GFX9-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX9-NEXT:    v_lshl_or_b32 v12, v43, 16, v12
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, s74, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v11, 0x300, v1
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_add_u32_e32 v1, 3, v1
; GFX9-NEXT:    v_perm_b32 v1, v1, v2, s4
; GFX9-NEXT:    v_add_u32_e32 v44, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s62
; GFX9-NEXT:    v_perm_b32 v1, s72, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v10, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s44
; GFX9-NEXT:    v_perm_b32 v1, s56, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v45, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s57
; GFX9-NEXT:    v_perm_b32 v1, s60, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v9, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s13
; GFX9-NEXT:    v_perm_b32 v1, s40, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v46, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s41
; GFX9-NEXT:    v_perm_b32 v1, s76, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v8, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s9
; GFX9-NEXT:    v_perm_b32 v1, s11, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v47, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s73
; GFX9-NEXT:    v_perm_b32 v1, s75, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v7, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s45
; GFX9-NEXT:    v_perm_b32 v1, s58, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v56, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s61
; GFX9-NEXT:    v_perm_b32 v1, s63, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v6, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s42
; GFX9-NEXT:    v_perm_b32 v1, s46, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v57, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s47
; GFX9-NEXT:    v_perm_b32 v1, s59, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v5, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s12
; GFX9-NEXT:    v_perm_b32 v1, s14, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v58, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s15
; GFX9-NEXT:    v_perm_b32 v1, s43, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v4, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s8
; GFX9-NEXT:    v_perm_b32 v1, s10, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v59, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s29
; GFX9-NEXT:    v_perm_b32 v1, s28, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v3, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    v_perm_b32 v1, s7, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v60, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s25
; GFX9-NEXT:    v_perm_b32 v1, s24, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v2, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s27
; GFX9-NEXT:    v_perm_b32 v1, s26, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v61, 0x300, v1
; GFX9-NEXT:    v_mov_b32_e32 v1, s21
; GFX9-NEXT:    v_perm_b32 v1, s20, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x300, v1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; GFX9-NEXT:    v_lshl_or_b32 v1, v32, 16, v1
; GFX9-NEXT:    v_and_b32_e32 v32, 0xffff, v54
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v32
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; GFX9-NEXT:    v_lshl_or_b32 v29, v30, 16, v29
; GFX9-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshl_or_b32 v26, v31, 16, v26
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v0, s18, v34, v0
; GFX9-NEXT:    v_add_u32_e32 v0, 0x300, v0
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX9-NEXT:    v_lshl_or_b32 v27, v28, 16, v27
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff, v38
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v2, v61, 16, v2
; GFX9-NEXT:    v_lshl_or_b32 v3, v60, 16, v3
; GFX9-NEXT:    v_lshl_or_b32 v4, v59, 16, v4
; GFX9-NEXT:    v_lshl_or_b32 v5, v58, 16, v5
; GFX9-NEXT:    v_lshl_or_b32 v6, v57, 16, v6
; GFX9-NEXT:    v_lshl_or_b32 v7, v56, 16, v7
; GFX9-NEXT:    v_lshl_or_b32 v8, v47, 16, v8
; GFX9-NEXT:    v_lshl_or_b32 v9, v46, 16, v9
; GFX9-NEXT:    v_lshl_or_b32 v10, v45, 16, v10
; GFX9-NEXT:    v_lshl_or_b32 v11, v44, 16, v11
; GFX9-NEXT:    v_lshl_or_b32 v28, v62, 16, v28
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_lshl_or_b32 v30, v31, 16, v30
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; GFX9-NEXT:    v_lshl_or_b32 v31, v32, 16, v31
; GFX9-NEXT:  .LBB97_5: ; %end
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v128i8_to_v64i16_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v133, off, s32 offset:312
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v166, off, s32 offset:308
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v161, off, s32 offset:304
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v180, off, s32 offset:300
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v129, off, s32 offset:296
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v165, off, s32 offset:292
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v178, off, s32 offset:288
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v182, off, s32 offset:284
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v102, off, s32 offset:280
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v160, off, s32 offset:276
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v163, off, s32 offset:272
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v181, off, s32 offset:268
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v132, off, s32 offset:264
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v162, off, s32 offset:260
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v150, off, s32 offset:256
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v179, off, s32 offset:252
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v101, off, s32 offset:248
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v149, off, s32 offset:244
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v146, off, s32 offset:240
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v177, off, s32 offset:236
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v87, off, s32 offset:232
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v144, off, s32 offset:228
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v128, off, s32 offset:224
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v167, off, s32 offset:220
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v82, off, s32 offset:216
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v131, off, s32 offset:212
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v148, off, s32 offset:208
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v176, off, s32 offset:204
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v64, off, s32 offset:200
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v117, off, s32 offset:196
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v119, off, s32 offset:192
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v164, off, s32 offset:188
; GFX11-TRUE16-NEXT:    s_clause 0x1f
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v85, off, s32 offset:184
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v118, off, s32 offset:180
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v115, off, s32 offset:176
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v151, off, s32 offset:172
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v55, off, s32 offset:168
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v113, off, s32 offset:164
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v100, off, s32 offset:160
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v147, off, s32 offset:156
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v52, off, s32 offset:152
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v98, off, s32 offset:148
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v81, off, s32 offset:144
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v135, off, s32 offset:140
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v49, off, s32 offset:136
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v84, off, s32 offset:132
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v112, off, s32 offset:128
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v145, off, s32 offset:124
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v36, off, s32 offset:120
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v69, off, s32 offset:116
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v80, off, s32 offset:112
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v130, off, s32 offset:108
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v51, off, s32 offset:104
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v71, off, s32 offset:100
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v67, off, s32 offset:96
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v116, off, s32 offset:92
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v35, off, s32 offset:88
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v66, off, s32 offset:84
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v54, off, s32 offset:80
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v103, off, s32 offset:76
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v34, off, s32 offset:72
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v53, off, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v48, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v97, off, s32 offset:60
; GFX11-TRUE16-NEXT:    s_clause 0xf
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v33, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v50, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v65, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v99, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v32, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v38, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v39, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v83, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v70, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v114, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v96, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v134, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v37, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v68, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_d16_b16 v86, off, s32
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s74, v30
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s10, v29
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v28
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v27
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s73, v26
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s13, v25
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v24
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v23
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s62, v22
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s11, v21
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v20
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v19
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s59, v18
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s5, v17
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s56, v16
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s57, v15
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s63, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s6, v13
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s12, v12
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s58, v11
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s72, v10
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s7, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s14, v8
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s15, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s60, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s8, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s61, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s4, v1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s9, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s75, 0
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:324
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:320
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-TRUE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB97_4
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, 0xc0c0004 :: v_dual_lshlrev_b32 v13, 8, v86
; GFX11-TRUE16-NEXT:    s_and_b32 s76, s43, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s77, s11, 8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s2, s3, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s22, s23, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s26, s27, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s0, s1, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s18, s19, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s20, s21, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v3.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s9, s4, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s24, s25, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s40, s8, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s14, s7, v8
; GFX11-TRUE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, s16, s17, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s12, s6, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v4.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s28, s29, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s76 :: v_dual_lshlrev_b32 v15, 8, v70
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v5.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s61, s45, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v6.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, s60, s15, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v7.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, s72, s58, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v9.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, s59, s44, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v68
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s41, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s89, s10, 8
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, s73, s42, v8
; GFX11-TRUE16-NEXT:    s_or_b32 s88, s88, s89
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v12
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s88
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s74, 0xff
; GFX11-TRUE16-NEXT:    s_and_b32 s77, s63, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s76, s57, 8
; GFX11-TRUE16-NEXT:    s_and_b32 s78, s56, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s79, s5, 8
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v114
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, s88, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v38
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v13, 8, v32
; GFX11-TRUE16-NEXT:    s_or_b32 s76, s77, s76
; GFX11-TRUE16-NEXT:    s_or_b32 s77, s78, s79
; GFX11-TRUE16-NEXT:    s_and_b32 s78, s62, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s79, s47, 8
; GFX11-TRUE16-NEXT:    s_and_b32 s88, s46, 0xff
; GFX11-TRUE16-NEXT:    s_lshl_b32 s89, s13, 8
; GFX11-TRUE16-NEXT:    s_or_b32 s78, s78, s79
; GFX11-TRUE16-NEXT:    s_or_b32 s79, s88, s89
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v8.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v14, v15
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xff, v83
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v16, 8, v33
; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s77
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xff, v50
; GFX11-TRUE16-NEXT:    s_pack_ll_b32_b16 s77, s78, s79
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 8, v39
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v13
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v134, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v8.l
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v15, v16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, v14, v17
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v53, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v99, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v66, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v97, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v71, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v103, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v69, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v116, v67, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v84, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v130, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v98, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v145, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v10.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v113, v55, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v135, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v8.l
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v147, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v118, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v117
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v24, 8, v64
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v151, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v164
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v119
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xff, v131
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v82
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v24
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xff, v149
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, v8, v25
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v176, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v26, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v144
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v26, 8, v87
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v128
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v167
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v101
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xff, v160
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v102
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, v8, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, v28, v29
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xff, v162
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v28, 8, v132
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v177, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v8.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v179
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v150
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, v10, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v133
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v40, 0xff, v182
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, v8, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xff, v166
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v10.l
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xff, v165
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v129
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v41, v29, v31
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v178
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v181, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v183, v8, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v30.l
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v8, s76
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, v40, v31
; GFX11-TRUE16-NEXT:    v_perm_b32 v31, v180, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v10, s77
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v30.h, v183.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v41.l
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB97_3
; GFX11-TRUE16-NEXT:  .LBB97_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v180
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v166
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v182
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v165
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v181
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v178, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s59, s59, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v160
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v179
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v162
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v129, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v1, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v177
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v102, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v149
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v132, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v167
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v144
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v176
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v101, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v131
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v164
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v87, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v117
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v151
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v118
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v82, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v147
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v113
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v85, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v135
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v55, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v98
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v145
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v55, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v130
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v52, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v69
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v116
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v71
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v67, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v103
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 3, v66
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v97
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v0, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v53
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v99
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v2, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v50
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 3, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 3, v38
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v39, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v134
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 3, v114
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v2, v32, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v3, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 3, v68
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v4, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s74, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v1
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s73, s73, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, v3, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_add_i32 s72, s72, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v2
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s73, s42, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s62, s47, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s59, s44, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s43, s43, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s72, s58, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s43, s11, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s56, s5, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s41, s41, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s12, s6, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s61, s61, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s41, s10, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, s46, s13, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s60, s15, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, s61, s45, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v65, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s28, s29, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s63, s63, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s9, s9, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s63, s57, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v48, 0x300, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v4
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, s9, s4, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s20, s21, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s14, s7, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v54, s40, s8, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v67, s26, s27, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v68, 0x300, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v0
; GFX11-TRUE16-NEXT:    v_perm_b32 v0, s22, s23, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v69, s2, s3, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v70, s18, s19, v1
; GFX11-TRUE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-TRUE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v66, 0x300, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, s24, s25, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v71, s16, s17, v1
; GFX11-TRUE16-NEXT:    v_perm_b32 v80, s0, s1, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v54, 0x300, v54
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v67, 0x300, v67
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v69
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v70
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v81, 0x300, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x300, v71
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x300, v80
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v69.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v70.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v81.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v67.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v68.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v54.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v66.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v65.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v53.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v50.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v48.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v39.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v37.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v38.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v32.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v33.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v34.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v35.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v51.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v36.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v49.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v52.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v55.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v85.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v64.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v82.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v87.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v101.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v28.h, v132.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v102.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v30.h, v129.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e64 v31.h, v133.l
; GFX11-TRUE16-NEXT:  .LBB97_3: ; %end
; GFX11-TRUE16-NEXT:    s_clause 0x1 ; 8-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:320
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:324
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-TRUE16-NEXT:  .LBB97_4:
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_branch .LBB97_2
;
; GFX11-FAKE16-LABEL: bitcast_v128i8_to_v64i16_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v166, off, s32 offset:312
; GFX11-FAKE16-NEXT:    scratch_load_u16 v180, off, s32 offset:308
; GFX11-FAKE16-NEXT:    scratch_load_u16 v167, off, s32 offset:304
; GFX11-FAKE16-NEXT:    scratch_load_u16 v182, off, s32 offset:300
; GFX11-FAKE16-NEXT:    scratch_load_u16 v150, off, s32 offset:296
; GFX11-FAKE16-NEXT:    scratch_load_u16 v178, off, s32 offset:292
; GFX11-FAKE16-NEXT:    scratch_load_u16 v165, off, s32 offset:288
; GFX11-FAKE16-NEXT:    scratch_load_u16 v181, off, s32 offset:284
; GFX11-FAKE16-NEXT:    scratch_load_u16 v161, off, s32 offset:280
; GFX11-FAKE16-NEXT:    scratch_load_u16 v176, off, s32 offset:276
; GFX11-FAKE16-NEXT:    scratch_load_u16 v163, off, s32 offset:272
; GFX11-FAKE16-NEXT:    scratch_load_u16 v179, off, s32 offset:268
; GFX11-FAKE16-NEXT:    scratch_load_u16 v135, off, s32 offset:264
; GFX11-FAKE16-NEXT:    scratch_load_u16 v151, off, s32 offset:260
; GFX11-FAKE16-NEXT:    scratch_load_u16 v164, off, s32 offset:256
; GFX11-FAKE16-NEXT:    scratch_load_u16 v177, off, s32 offset:252
; GFX11-FAKE16-NEXT:    scratch_load_u16 v133, off, s32 offset:248
; GFX11-FAKE16-NEXT:    scratch_load_u16 v149, off, s32 offset:244
; GFX11-FAKE16-NEXT:    scratch_load_u16 v145, off, s32 offset:240
; GFX11-FAKE16-NEXT:    scratch_load_u16 v162, off, s32 offset:236
; GFX11-FAKE16-NEXT:    scratch_load_u16 v144, off, s32 offset:232
; GFX11-FAKE16-NEXT:    scratch_load_u16 v148, off, s32 offset:228
; GFX11-FAKE16-NEXT:    scratch_load_u16 v147, off, s32 offset:224
; GFX11-FAKE16-NEXT:    scratch_load_u16 v160, off, s32 offset:220
; GFX11-FAKE16-NEXT:    scratch_load_u16 v114, off, s32 offset:216
; GFX11-FAKE16-NEXT:    scratch_load_u16 v131, off, s32 offset:212
; GFX11-FAKE16-NEXT:    scratch_load_u16 v128, off, s32 offset:208
; GFX11-FAKE16-NEXT:    scratch_load_u16 v146, off, s32 offset:204
; GFX11-FAKE16-NEXT:    scratch_load_u16 v117, off, s32 offset:200
; GFX11-FAKE16-NEXT:    scratch_load_u16 v130, off, s32 offset:196
; GFX11-FAKE16-NEXT:    scratch_load_u16 v119, off, s32 offset:192
; GFX11-FAKE16-NEXT:    scratch_load_u16 v134, off, s32 offset:188
; GFX11-FAKE16-NEXT:    s_clause 0x1f
; GFX11-FAKE16-NEXT:    scratch_load_u16 v99, off, s32 offset:184
; GFX11-FAKE16-NEXT:    scratch_load_u16 v115, off, s32 offset:180
; GFX11-FAKE16-NEXT:    scratch_load_u16 v129, off, s32 offset:176
; GFX11-FAKE16-NEXT:    scratch_load_u16 v132, off, s32 offset:172
; GFX11-FAKE16-NEXT:    scratch_load_u16 v97, off, s32 offset:168
; GFX11-FAKE16-NEXT:    scratch_load_u16 v113, off, s32 offset:164
; GFX11-FAKE16-NEXT:    scratch_load_u16 v101, off, s32 offset:160
; GFX11-FAKE16-NEXT:    scratch_load_u16 v118, off, s32 offset:156
; GFX11-FAKE16-NEXT:    scratch_load_u16 v100, off, s32 offset:152
; GFX11-FAKE16-NEXT:    scratch_load_u16 v112, off, s32 offset:148
; GFX11-FAKE16-NEXT:    scratch_load_u16 v103, off, s32 offset:144
; GFX11-FAKE16-NEXT:    scratch_load_u16 v116, off, s32 offset:140
; GFX11-FAKE16-NEXT:    scratch_load_u16 v70, off, s32 offset:136
; GFX11-FAKE16-NEXT:    scratch_load_u16 v87, off, s32 offset:132
; GFX11-FAKE16-NEXT:    scratch_load_u16 v84, off, s32 offset:128
; GFX11-FAKE16-NEXT:    scratch_load_u16 v102, off, s32 offset:124
; GFX11-FAKE16-NEXT:    scratch_load_u16 v81, off, s32 offset:120
; GFX11-FAKE16-NEXT:    scratch_load_u16 v86, off, s32 offset:116
; GFX11-FAKE16-NEXT:    scratch_load_u16 v83, off, s32 offset:112
; GFX11-FAKE16-NEXT:    scratch_load_u16 v98, off, s32 offset:108
; GFX11-FAKE16-NEXT:    scratch_load_u16 v54, off, s32 offset:104
; GFX11-FAKE16-NEXT:    scratch_load_u16 v71, off, s32 offset:100
; GFX11-FAKE16-NEXT:    scratch_load_u16 v85, off, s32 offset:96
; GFX11-FAKE16-NEXT:    scratch_load_u16 v96, off, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_load_u16 v52, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_u16 v69, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_u16 v64, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_u16 v82, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_u16 v55, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_u16 v68, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_u16 v66, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_u16 v80, off, s32 offset:60
; GFX11-FAKE16-NEXT:    s_clause 0xf
; GFX11-FAKE16-NEXT:    scratch_load_u16 v34, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_u16 v50, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_u16 v39, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_u16 v65, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_u16 v36, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_u16 v49, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_u16 v37, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_u16 v53, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32 offset:316
; GFX11-FAKE16-NEXT:    scratch_load_u16 v32, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_u16 v48, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_u16 v51, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_u16 v67, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_u16 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_u16 v38, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_u16 v35, off, s32
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s74, v30
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s61, v29
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s63, v28
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s72, v27
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s73, v26
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v25
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s56, v24
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s59, v23
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s62, v22
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v21
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v20
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s57, v19
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s60, v18
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s13, v17
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v16
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v15
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s58, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s10, v13
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s14, v12
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v11
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v10
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s9, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s12, v8
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s15, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s5, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s7, v4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s8, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s11, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s4, v1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s6, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s75, 0
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:320 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(7)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v31
; GFX11-FAKE16-NEXT:    s_and_b32 s76, vcc_lo, exec_lo
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB97_4
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s74, 0xff
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v12, 8, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xff, v48
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v14, 8, v32
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s0, s1, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s16, s17, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s20, s21, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s2, s3, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s18, s19, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s22, s23, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, s24, s25, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s28, s29, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v3, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s11, s8, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v3, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s6, s4, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v6, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s7, s5, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v67, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v80, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v6
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v68, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v10, 16, v11
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v35
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s26, s27, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, s42, s15, v8
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s44, s41, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v69, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v7, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s12, s9, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v116, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_lshl_b32 s77, s43, 8
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v118, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, s14, s10, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, s76, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v38
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s46, 0xff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v7, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v11, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v13, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v53
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v9, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xff, v49
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v10, 8, v36
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v11, 16, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v14, v15
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xff, v50
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v15, 8, v34
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v65, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, v14, v15
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v82, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v115, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v130
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v9, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v96, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v71, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v98, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v86, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v117
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v144
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v102, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, s60, s57, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v19, 16, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v87, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v135
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v150
; GFX11-FAKE16-NEXT:    s_lshl_b32 s77, s47, 8
; GFX11-FAKE16-NEXT:    s_and_b32 s78, s40, 0xff
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v11, 16, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v112, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, s76, 16, v10
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v132, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_and_b32 s76, s58, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s79, s13, 8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v11, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v113, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_or_b32 s76, s76, s77
; GFX11-FAKE16-NEXT:    s_or_b32 s77, s78, s79
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, s73, s72, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v11, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v23, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v134
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v119
; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s76, s76, s77
; GFX11-FAKE16-NEXT:    s_and_b32 s77, s62, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s78, s59, 8
; GFX11-FAKE16-NEXT:    s_lshl_b32 s79, s45, 8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v24, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xff, v131
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 8, v114
; GFX11-FAKE16-NEXT:    s_or_b32 s77, s77, s78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_and_b32 s78, s56, 0xff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v25, v24, v25
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v146, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_or_b32 s78, s78, s79
; GFX11-FAKE16-NEXT:    s_and_b32 s79, s63, 0xff
; GFX11-FAKE16-NEXT:    s_lshl_b32 s88, s61, 8
; GFX11-FAKE16-NEXT:    s_pack_ll_b32_b16 s77, s77, s78
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v160
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v147
; GFX11-FAKE16-NEXT:    s_or_b32 s78, s79, s88
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v26
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v148
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v26, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xff, v149
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 8, v133
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v27, v26, v27
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v162, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v177
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v164
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v27, 16, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v151
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v11, v28, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xff, v176
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 8, v161
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v29, v28, v29
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v179, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff, v28
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v10, 0xff, v181
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v11, 8, v165
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v29, 16, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xff, v178
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, v10, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v11, 0xff, v180
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v30, v30, v31
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 8, v166
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v183, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v10, s77
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v31, v11, v31
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v182, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v183
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v40, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, s78, 16, v8
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v8, s76
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v40
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s75
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB97_3
; GFX11-FAKE16-NEXT:  .LBB97_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v182
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v180
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v178
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v181
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v179
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 3, v176
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v165, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s74, s74, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v27, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v177
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v31, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v3, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v4, v163, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v5, v161, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v28, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v30, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v151
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v150, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v151, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v162
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v149
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v148
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v29, 0x300, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v133, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v144, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v132
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v135, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v146
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v160
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v26, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v131
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v144, 0x300, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v134
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v114, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s62, s62, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v23, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v129, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v119, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v25, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v115
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v22, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v115, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v118
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v130
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v24, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v113
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v112
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v101, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v102
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v114, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v116
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v21, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v87
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v96
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v103, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v100, 0x300, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v70, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v18, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v85, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v17, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v98
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v86
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v20, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v71
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v71, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v82
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v83, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v68
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v19, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v70, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v69
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v80
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v54, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v65
; GFX11-FAKE16-NEXT:    s_add_i32 s40, s40, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v52, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v4, v55, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v39, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 3, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v55, 0x300, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 3, v53
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 3, v49
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v4, v51, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v16, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v36, 0x300, v0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 3, v38
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x300, v2
; GFX11-FAKE16-NEXT:    v_mov_b32_e32 v2, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v34, 0x300, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 3, v48
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s74, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s73, s73, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v35, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s62, s59, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v32, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_add_i32 s63, s63, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s12, s12, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s73, s72, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s40, s13, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v32, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s63, s61, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s56, s56, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s60, s60, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v39, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s12, s9, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s46, s46, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s58, s58, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s11, s11, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s24, s24, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s0, s0, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v33, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x300, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s56, s45, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s60, s57, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s46, s43, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s58, s47, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x300, v0
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s11, s8, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v50, s24, s25, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v65, s0, s1, v2
; GFX11-FAKE16-NEXT:    s_add_i32 s44, s44, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s14, s14, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s42, s42, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s26, s26, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s2, s2, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v37, 0x300, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v38, 0x300, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s44, s41, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, s14, s10, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s42, s15, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v51, 0x300, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v50, 0x300, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, s26, s27, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v67, s2, s3, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v65, 0x300, v65
; GFX11-FAKE16-NEXT:    s_add_i32 s7, s7, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s28, s28, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s16, s16, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v48, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v49, 0x300, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x300, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s7, s5, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, s28, s29, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v64, s16, s17, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v67, 0x300, v67
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff, v65
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v69, 0x300, v0
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v50
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v38
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v33
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; GFX11-FAKE16-NEXT:    s_add_i32 s6, s6, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s20, s20, 3
; GFX11-FAKE16-NEXT:    s_add_i32 s18, s18, 3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, s6, s4, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v53, 0x300, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x300, v4
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, s20, s21, v2
; GFX11-FAKE16-NEXT:    v_perm_b32 v66, s18, s19, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v64, 0x300, v64
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v67, 16, v65
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v69, 16, v50
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v39, 16, v38
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v37, 16, v9
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v35, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v32, 16, v36
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v34, 16, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v54
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v71
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v17
; GFX11-FAKE16-NEXT:    s_add_i32 s22, s22, 3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x300, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, s22, s23, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v66, 0x300, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v64
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v68, 0x300, v1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v51
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v55, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v52, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v99
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v115
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x300, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v66, 16, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v68
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v5, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v53, 16, v51
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v100, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v97, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v25, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v32, 0xffff, v135
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v33, 0xffff, v151
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v34, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v35, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v2, 16, v64
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v6, 16, v65
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v49, 16, v48
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v10, 16, v50
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v11, 16, v51
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v15, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v38
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v70, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v114, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v144, 16, v32
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v133, 16, v33
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v150, 16, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v30, 16, v34
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v35
; GFX11-FAKE16-NEXT:  .LBB97_3: ; %end
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:320 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
; GFX11-FAKE16-NEXT:  .LBB97_4:
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-FAKE16-NEXT:    s_branch .LBB97_2
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <128 x i8> %a, splat (i8 3)
  %a2 = bitcast <128 x i8> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <128 x i8> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) {
; SI-LABEL: bitcast_v64i16_to_v128i8:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v17
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v23
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v2
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v29
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v4
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v8
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v10
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v12
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v58
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v14
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v57
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v16
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v56
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v18
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v47
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v20
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v46
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v22
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v45
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v24
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v44
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v26
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v43
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v28
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v42
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v30
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v41
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v40
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v60
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
; SI-NEXT:    buffer_store_dword v50, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v52, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; kill: killed $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB98_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr49
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v41, v1, v31
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    v_or_b32_e32 v36, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v47, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v44, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v5
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v52, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v6
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v37, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v7
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v35, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v34, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v9
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v33, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v10
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v32, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v11
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v31, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v12
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v63, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v62, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v61, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v59, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v16
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v58, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v17
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v57, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v18
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v56, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v46, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v20
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v45, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v42, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v36, v41, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v22
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v43, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v36, v41, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v23
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v55, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v36, v41, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v24
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v40, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v44, v47, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v25
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v53, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v44, v47, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v26
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v54, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v44, v47, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v27
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v50, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v37, v52, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v28
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v51, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v37, v52, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v29
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v3, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v37, v52, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v30
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v48, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v34, v35, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v39
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v2, v1, v2
; SI-NEXT:    v_alignbit_b32 v1, v34, v35, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v60
; SI-NEXT:    v_or_b32_e32 v38, v1, v4
; SI-NEXT:    v_alignbit_b32 v1, v34, v35, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v32, v33, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v32, v33, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v32, v33, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v63, v31, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v63, v31, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v63, v31, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v61, v62, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v61, v62, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v61, v62, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v59, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v59, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v59, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v57, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v57, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v57, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v43, v42, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v43, v42, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v43, v42, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v40, v55, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v40, v55, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v40, v55, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v53, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v53, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v53, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v51, v50, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v51, v50, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v51, v50, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v48, v3, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v48, v3, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v48, v3, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v38, v2, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v38, v2, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v38, v2, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v44
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v37
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v34
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v32
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v63
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v61
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v58
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v56
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v45
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v40
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v54
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v51
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v48
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_bfe_u32 v1, v1, 8, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:  .LBB98_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB98_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v31, vcc, 3, v39
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_mov_b32 s6, 0x30000
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    v_add_i32_e32 v32, vcc, 3, v60
; SI-NEXT:    v_and_b32_e32 v32, 0xffff, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v32, v33, v32
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v39, vcc, 0x30000, v31
; SI-NEXT:    v_add_i32_e32 v38, vcc, s6, v32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v29, v33, v29
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v30, v33, v30
; SI-NEXT:    v_add_i32_e32 v48, vcc, s6, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v30, v27
; SI-NEXT:    v_add_i32_e32 v50, vcc, s6, v27
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    v_add_i32_e32 v51, vcc, s6, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v27, v25
; SI-NEXT:    v_add_i32_e32 v53, vcc, s6, v25
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    v_add_i32_e32 v54, vcc, s6, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v25, v23
; SI-NEXT:    v_add_i32_e32 v55, vcc, s6, v23
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    v_add_i32_e32 v40, vcc, s6, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v23, v21
; SI-NEXT:    v_add_i32_e32 v42, vcc, s6, v21
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    v_add_i32_e32 v43, vcc, s6, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v21, v19
; SI-NEXT:    v_add_i32_e32 v46, vcc, s6, v19
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v19, v20, v19
; SI-NEXT:    v_add_i32_e32 v45, vcc, s6, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v19, v17
; SI-NEXT:    v_add_i32_e32 v57, vcc, s6, v17
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    v_add_i32_e32 v56, vcc, s6, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v17, v15
; SI-NEXT:    v_add_i32_e32 v59, vcc, s6, v15
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v15, v16, v15
; SI-NEXT:    v_add_i32_e32 v58, vcc, s6, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v15, v13
; SI-NEXT:    v_add_i32_e32 v62, vcc, s6, v13
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v13, v14, v13
; SI-NEXT:    v_add_i32_e32 v61, vcc, s6, v13
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v13, v11
; SI-NEXT:    v_add_i32_e32 v31, vcc, s6, v11
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    v_add_i32_e32 v63, vcc, s6, v11
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v9, v11, v9
; SI-NEXT:    v_add_i32_e32 v33, vcc, s6, v9
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    v_add_i32_e32 v32, vcc, s6, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v7, v9, v7
; SI-NEXT:    v_add_i32_e32 v35, vcc, s6, v7
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v8
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_add_i32_e32 v34, vcc, s6, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v5, v7, v5
; SI-NEXT:    v_add_i32_e32 v52, vcc, s6, v5
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_add_i32_e32 v37, vcc, s6, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v3, v5, v3
; SI-NEXT:    v_add_i32_e32 v47, vcc, s6, v3
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    v_add_i32_e32 v44, vcc, s6, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v3, v1
; SI-NEXT:    v_add_i32_e32 v41, vcc, s6, v1
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    v_add_i32_e32 v36, vcc, s6, v1
; SI-NEXT:    v_alignbit_b32 v1, v36, v41, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v41, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v41, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v47, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v47, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v47, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v37, v52, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v37, v52, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v37, v52, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v34, v35, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v34, v35, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v34, v35, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v32, v33, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v32, v33, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v32, v33, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v63, v31, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v63, v31, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v63, v31, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v61, v62, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v61, v62, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v61, v62, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v59, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v59, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v58, v59, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v57, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v57, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v56, v57, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v45, v46, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v43, v42, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v43, v42, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v43, v42, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v40, v55, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v40, v55, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v40, v55, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v53, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v53, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v54, v53, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v51, v50, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v51, v50, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v51, v50, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v48, v29, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v48, v29, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v48, v29, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v38, v39, 24
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v38, v39, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v38, v39, 8
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v44
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v44
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v44
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v37
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v37
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v37
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v34
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v34
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v34
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v32
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v32
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v32
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v63
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v63
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v63
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v61
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v61
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v61
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v58
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v58
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v58
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v56
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v56
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v56
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v45
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v45
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v45
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v40
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v40
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v40
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v54
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v54
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v54
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v51
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v51
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v51
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v48
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v48
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v48
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 24, v38
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v38
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v38
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:  .LBB98_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v41
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v36
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 4, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v47
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 8, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v44
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 12, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v52
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 16, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v37
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 20, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v35
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 24, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v34
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 28, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v33
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 32, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v32
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 36, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v31
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 40, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v63
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v62
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v61
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 52, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v59
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 56, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v58
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v57
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v56
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x44, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v46
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x48, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v45
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x4c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v42
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x50, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v43
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x54, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v55
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x58, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v40
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x5c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v53
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x60, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v54
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x64, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v50
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x68, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v51
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x6c, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x70, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v48
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x74, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v1
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, 0x78, v0
; SI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xff, v38
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 24, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v128i8:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:8
; VI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v36, off, s[0:3], s32
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v15
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v13
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v11
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v9
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v7
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v5
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v30
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v29
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v26
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v63, 16, v16
; VI-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; VI-NEXT:    v_lshrrev_b32_e32 v60, 16, v12
; VI-NEXT:    v_lshrrev_b32_e32 v61, 16, v10
; VI-NEXT:    v_lshrrev_b32_e32 v62, 16, v8
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v6
; VI-NEXT:    v_lshrrev_b32_e32 v46, 16, v4
; VI-NEXT:    v_lshrrev_b32_e32 v47, 16, v2
; VI-NEXT:    v_lshrrev_b32_e32 v45, 16, v28
; VI-NEXT:    v_lshrrev_b32_e32 v35, 16, v27
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v59, 16, v25
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v24
; VI-NEXT:    v_lshrrev_b32_e32 v50, 16, v23
; VI-NEXT:    v_lshrrev_b32_e32 v51, 16, v22
; VI-NEXT:    v_lshrrev_b32_e32 v52, 16, v21
; VI-NEXT:    v_lshrrev_b32_e32 v55, 16, v20
; VI-NEXT:    v_lshrrev_b32_e32 v57, 16, v18
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v19
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v36
; VI-NEXT:    v_lshrrev_b32_e32 v53, 16, v37
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 16, v17
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr58
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr40
; VI-NEXT:    ; implicit-def: $vgpr48
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr32
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr49
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr38
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr56
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; kill: killed $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr41
; VI-NEXT:    ; implicit-def: $vgpr31
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr44
; VI-NEXT:    ; implicit-def: $vgpr43
; VI-NEXT:    ; implicit-def: $vgpr42
; VI-NEXT:    ; implicit-def: $vgpr39
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    ; implicit-def: $vgpr54
; VI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB98_2
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    v_lshrrev_b32_e32 v31, 24, v16
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v16
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v15
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 24, v14
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v14
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v13
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 24, v12
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v12
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v11
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 24, v10
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v10
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v9
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 24, v8
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v8
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v7
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 24, v6
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v6
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v5
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[31:32], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[31:32], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[31:32], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[31:32], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v31, v7
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v10
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v11
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v12
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v13
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v14
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v15
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v16
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v9, v8
; VI-NEXT:    v_lshrrev_b64 v[7:8], 24, v[7:8]
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v5
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v7, v6
; VI-NEXT:    v_lshrrev_b64 v[5:6], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v5, 24, v4
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v5, v3
; VI-NEXT:    v_lshrrev_b64 v[31:32], 24, v[3:4]
; VI-NEXT:    v_lshrrev_b32_e32 v3, 24, v2
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v3, 8, v2
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v3, 8, v1
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v3, v1
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    v_lshrrev_b32_e32 v1, 24, v37
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v37
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v36
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v36
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v37
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v3, v2
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[36:37]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 24, v30
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v30
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v29
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v29
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v30
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 24, v28
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v28
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v27
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v27
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v28
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 24, v26
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v26
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v25
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v25
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v26
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v24
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v23
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v23
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v24
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v22
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v21
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v21
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v22
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v20
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v19
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v20
; VI-NEXT:    v_lshrrev_b64 v[42:43], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v17
; VI-NEXT:    v_lshrrev_b64 v[43:44], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v18
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v5, v4
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v45
; VI-NEXT:    v_lshrrev_b64 v[44:45], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v56, 24, v24
; VI-NEXT:    v_lshrrev_b32_e32 v38, 24, v22
; VI-NEXT:    v_lshrrev_b32_e32 v49, 24, v20
; VI-NEXT:    v_lshrrev_b32_e32 v32, 8, v19
; VI-NEXT:    v_lshrrev_b32_e32 v48, 24, v18
; VI-NEXT:    v_lshrrev_b32_e32 v40, 8, v18
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v17
; VI-NEXT:    v_mov_b32_e32 v45, v1
; VI-NEXT:    ; implicit-def: $vgpr1
; VI-NEXT:    ; implicit-def: $vgpr3
; VI-NEXT:    ; implicit-def: $vgpr5
; VI-NEXT:    ; implicit-def: $vgpr7
; VI-NEXT:    ; implicit-def: $vgpr9
; VI-NEXT:    ; implicit-def: $vgpr11
; VI-NEXT:    ; implicit-def: $vgpr13
; VI-NEXT:    ; implicit-def: $vgpr15
; VI-NEXT:    ; implicit-def: $vgpr17
; VI-NEXT:    ; implicit-def: $vgpr19
; VI-NEXT:    ; implicit-def: $vgpr21
; VI-NEXT:    ; implicit-def: $vgpr23
; VI-NEXT:    ; implicit-def: $vgpr25
; VI-NEXT:    ; implicit-def: $vgpr27
; VI-NEXT:    ; implicit-def: $vgpr29
; VI-NEXT:    ; implicit-def: $vgpr37
; VI-NEXT:  .LBB98_2: ; %Flow
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB98_4
; VI-NEXT:  ; %bb.3: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v31, 3
; VI-NEXT:    v_add_u16_sdwa v55, v18, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v32, 3, v18
; VI-NEXT:    v_lshlrev_b32_e32 v18, 16, v55
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v18, v32, v18
; VI-NEXT:    v_add_u16_e32 v32, 3, v17
; VI-NEXT:    v_add_u16_sdwa v17, v17, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_add_u16_sdwa v54, v20, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_add_u16_e32 v32, 3, v20
; VI-NEXT:    v_lshlrev_b32_e32 v20, 16, v54
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v20, v32, v20
; VI-NEXT:    v_add_u16_e32 v32, 3, v19
; VI-NEXT:    v_add_u16_sdwa v19, v19, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_add_u16_sdwa v57, v22, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v19, v32, v19
; VI-NEXT:    v_add_u16_e32 v32, 3, v22
; VI-NEXT:    v_lshlrev_b32_e32 v22, 16, v57
; VI-NEXT:    v_add_u16_sdwa v52, v21, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v22, v32, v22
; VI-NEXT:    v_add_u16_e32 v32, 3, v21
; VI-NEXT:    v_lshlrev_b32_e32 v21, 16, v52
; VI-NEXT:    v_add_u16_sdwa v56, v24, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v21, v32, v21
; VI-NEXT:    v_add_u16_e32 v32, 3, v24
; VI-NEXT:    v_lshlrev_b32_e32 v24, 16, v56
; VI-NEXT:    v_add_u16_sdwa v50, v23, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v24, v32, v24
; VI-NEXT:    v_add_u16_e32 v32, 3, v23
; VI-NEXT:    v_lshlrev_b32_e32 v23, 16, v50
; VI-NEXT:    v_add_u16_sdwa v49, v26, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v23, v32, v23
; VI-NEXT:    v_add_u16_e32 v32, 3, v26
; VI-NEXT:    v_lshlrev_b32_e32 v26, 16, v49
; VI-NEXT:    v_add_u16_sdwa v59, v25, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v26, v32, v26
; VI-NEXT:    v_add_u16_e32 v32, 3, v25
; VI-NEXT:    v_lshlrev_b32_e32 v25, 16, v59
; VI-NEXT:    v_add_u16_sdwa v38, v28, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v25, v32, v25
; VI-NEXT:    v_add_u16_e32 v32, 3, v28
; VI-NEXT:    v_lshlrev_b32_e32 v28, 16, v38
; VI-NEXT:    v_add_u16_sdwa v35, v27, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v28, v32, v28
; VI-NEXT:    v_add_u16_e32 v32, 3, v27
; VI-NEXT:    v_lshlrev_b32_e32 v27, 16, v35
; VI-NEXT:    v_add_u16_sdwa v48, v30, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v27, v32, v27
; VI-NEXT:    v_add_u16_e32 v32, 3, v30
; VI-NEXT:    v_add_u16_e32 v33, 3, v29
; VI-NEXT:    v_add_u16_sdwa v34, v29, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v48
; VI-NEXT:    v_or_b32_e32 v30, v32, v29
; VI-NEXT:    v_lshlrev_b32_e32 v29, 16, v34
; VI-NEXT:    v_add_u16_sdwa v53, v37, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v29, v33, v29
; VI-NEXT:    v_add_u16_e32 v34, 3, v37
; VI-NEXT:    v_add_u16_sdwa v33, v36, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v53
; VI-NEXT:    v_add_u16_e32 v39, 3, v36
; VI-NEXT:    v_or_b32_e32 v37, v34, v32
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; VI-NEXT:    v_add_u16_sdwa v47, v2, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v36, v39, v32
; VI-NEXT:    v_add_u16_e32 v33, 3, v2
; VI-NEXT:    v_add_u16_e32 v34, 3, v1
; VI-NEXT:    v_add_u16_sdwa v32, v1, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v47
; VI-NEXT:    v_or_b32_e32 v2, v33, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v32
; VI-NEXT:    v_add_u16_sdwa v46, v4, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v1, v34, v1
; VI-NEXT:    v_add_u16_e32 v33, 3, v4
; VI-NEXT:    v_add_u16_e32 v34, 3, v3
; VI-NEXT:    v_add_u16_sdwa v32, v3, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v46
; VI-NEXT:    v_or_b32_e32 v4, v33, v3
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v32
; VI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v3, v34, v3
; VI-NEXT:    v_add_u16_sdwa v34, v6, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; VI-NEXT:    v_add_u16_e32 v33, 3, v6
; VI-NEXT:    v_add_u16_e32 v39, 3, v5
; VI-NEXT:    v_add_u16_sdwa v32, v5, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v34
; VI-NEXT:    v_or_b32_e32 v6, v33, v5
; VI-NEXT:    v_lshlrev_b32_e32 v5, 16, v32
; VI-NEXT:    v_add_u16_sdwa v62, v8, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v5, v39, v5
; VI-NEXT:    v_add_u16_e32 v39, 3, v8
; VI-NEXT:    v_add_u16_e32 v33, 3, v7
; VI-NEXT:    v_add_u16_sdwa v32, v7, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v62
; VI-NEXT:    v_or_b32_e32 v8, v39, v7
; VI-NEXT:    v_lshlrev_b32_e32 v7, 16, v32
; VI-NEXT:    v_add_u16_sdwa v61, v10, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v7, v33, v7
; VI-NEXT:    v_add_u16_e32 v33, 3, v10
; VI-NEXT:    v_add_u16_e32 v39, 3, v9
; VI-NEXT:    v_add_u16_sdwa v32, v9, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v61
; VI-NEXT:    v_or_b32_e32 v10, v33, v9
; VI-NEXT:    v_lshlrev_b32_e32 v9, 16, v32
; VI-NEXT:    v_add_u16_sdwa v60, v12, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v9, v39, v9
; VI-NEXT:    v_add_u16_e32 v33, 3, v12
; VI-NEXT:    v_add_u16_e32 v39, 3, v11
; VI-NEXT:    v_add_u16_sdwa v32, v11, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v60
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v12, v33, v11
; VI-NEXT:    v_lshlrev_b32_e32 v11, 16, v32
; VI-NEXT:    v_add_u16_sdwa v33, v14, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v11, v39, v11
; VI-NEXT:    v_add_u16_e32 v39, 3, v14
; VI-NEXT:    v_add_u16_e32 v40, 3, v13
; VI-NEXT:    v_add_u16_sdwa v32, v13, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v33
; VI-NEXT:    v_add_u16_sdwa v63, v16, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v39, v13
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; VI-NEXT:    v_lshlrev_b32_e32 v13, 16, v32
; VI-NEXT:    v_add_u16_sdwa v31, v15, v31 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_add_u16_e32 v32, 3, v15
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v63
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v16, v16, v15
; VI-NEXT:    v_lshlrev_b32_e32 v15, 16, v31
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v15, v32, v15
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v16
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v31, 8, v15
; VI-NEXT:    v_lshrrev_b64 v[15:16], 24, v[15:16]
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; VI-NEXT:    v_or_b32_e32 v13, v40, v13
; VI-NEXT:    v_lshrrev_b32_e32 v15, 8, v14
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v15, 8, v13
; VI-NEXT:    v_lshrrev_b64 v[13:14], 24, v[13:14]
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v13, 8, v12
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v13, 8, v11
; VI-NEXT:    v_lshrrev_b64 v[11:12], 24, v[11:12]
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v11, 8, v10
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v11, 8, v9
; VI-NEXT:    v_lshrrev_b64 v[9:10], 24, v[9:10]
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v9, 8, v8
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v9, 8, v7
; VI-NEXT:    v_lshrrev_b64 v[7:8], 24, v[7:8]
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v6
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v7, 8, v5
; VI-NEXT:    v_lshrrev_b64 v[5:6], 24, v[5:6]
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v5, 8, v4
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
; VI-NEXT:    v_lshrrev_b64 v[31:32], 24, v[3:4]
; VI-NEXT:    v_lshrrev_b32_e32 v3, 8, v2
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v3, 8, v1
; VI-NEXT:    v_lshrrev_b64 v[41:42], 24, v[1:2]
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v37
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v36
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[36:37]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v30
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v29
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[29:30]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v28
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v27
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[27:28]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v26
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v25
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[1:2], 24, v[25:26]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v24
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v23
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v22
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v21
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b32_e32 v1, 8, v20
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v63, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v33, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v60, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v61, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v62, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v34, 8, 8
; VI-NEXT:    v_lshrrev_b64 v[42:43], 24, v[21:22]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v46, 8, 8
; VI-NEXT:    v_lshrrev_b64 v[43:44], 24, v[19:20]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v47, 8, 8
; VI-NEXT:    v_lshrrev_b64 v[44:45], 24, v[17:18]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v53, 8, 8
; VI-NEXT:    v_mov_b32_e32 v45, v38
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v48, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v45, 8, 8
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; VI-NEXT:    v_bfe_u32 v1, v49, 8, 8
; VI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; VI-NEXT:    v_lshrrev_b64 v[39:40], 24, v[23:24]
; VI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; VI-NEXT:    v_mov_b32_e32 v1, v57
; VI-NEXT:    v_mov_b32_e32 v57, v55
; VI-NEXT:    v_lshrrev_b32_e32 v32, 8, v19
; VI-NEXT:    v_lshrrev_b32_e32 v40, 8, v18
; VI-NEXT:    v_lshrrev_b32_e32 v58, 8, v17
; VI-NEXT:    v_bfe_u32 v56, v56, 8, 8
; VI-NEXT:    v_mov_b32_e32 v51, v1
; VI-NEXT:    v_bfe_u32 v38, v1, 8, 8
; VI-NEXT:    v_mov_b32_e32 v55, v54
; VI-NEXT:    v_bfe_u32 v49, v54, 8, 8
; VI-NEXT:    v_bfe_u32 v48, v57, 8, 8
; VI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; VI-NEXT:  .LBB98_4: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b32 s4, 0xc0c0004
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v1, v1, v41, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v47, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 4, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(2)
; VI-NEXT:    v_perm_b32 v1, v1, v31, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 8, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v46, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 12, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 16, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v34, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 20, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 24, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v62, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 28, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 32, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v61, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 36, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 40, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v60, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v33, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 52, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v3, v2, s4
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 56, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v2, v1, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v63, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v44, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v58, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; VI-NEXT:    v_perm_b32 v2, v57, v48, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v40, s4
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v1, v1, v43, s4
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v32, s4
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    v_perm_b32 v2, v55, v49, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    v_perm_b32 v2, v52, v42, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x50, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    v_perm_b32 v2, v51, v38, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x54, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    v_perm_b32 v2, v50, v39, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x58, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v56, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x5c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v59, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x60, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x64, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v35, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x68, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v45, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x6c, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x70, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x74, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_perm_b32 v2, v2, v3, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x78, v0
; VI-NEXT:    buffer_store_dword v1, v2, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v1, v1, v2, s4
; VI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_perm_b32 v2, v53, v2, s4
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v1, v1, v2
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64i16_to_v128i8:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:8
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr56
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    ; implicit-def: $vgpr63
; GFX9-NEXT:    ; implicit-def: $vgpr45
; GFX9-NEXT:    ; implicit-def: $vgpr44
; GFX9-NEXT:    ; implicit-def: $vgpr62
; GFX9-NEXT:    ; implicit-def: $vgpr61
; GFX9-NEXT:    ; implicit-def: $vgpr60
; GFX9-NEXT:    ; kill: killed $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr43
; GFX9-NEXT:    ; implicit-def: $vgpr59
; GFX9-NEXT:    ; implicit-def: $vgpr37
; GFX9-NEXT:    ; implicit-def: $vgpr58
; GFX9-NEXT:    ; implicit-def: $vgpr38
; GFX9-NEXT:    ; implicit-def: $vgpr36
; GFX9-NEXT:    ; implicit-def: $vgpr57
; GFX9-NEXT:    ; implicit-def: $vgpr39
; GFX9-NEXT:    ; implicit-def: $vgpr35
; GFX9-NEXT:    ; implicit-def: $vgpr34
; GFX9-NEXT:    ; implicit-def: $vgpr47
; GFX9-NEXT:    ; implicit-def: $vgpr46
; GFX9-NEXT:    ; kill: killed $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr42
; GFX9-NEXT:    ; implicit-def: $vgpr41
; GFX9-NEXT:    ; implicit-def: $vgpr40
; GFX9-NEXT:    ; implicit-def: $vgpr52
; GFX9-NEXT:    ; implicit-def: $vgpr51
; GFX9-NEXT:    ; implicit-def: $vgpr50
; GFX9-NEXT:    ; implicit-def: $vgpr49
; GFX9-NEXT:    ; implicit-def: $vgpr48
; GFX9-NEXT:    ; implicit-def: $vgpr54
; GFX9-NEXT:    ; implicit-def: $vgpr53
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr55
; GFX9-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(29)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v33
; GFX9-NEXT:    ; implicit-def: $vgpr33
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB98_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(38)
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v31
; GFX9-NEXT:    v_lshrrev_b64 v[48:49], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v30
; GFX9-NEXT:    v_lshrrev_b64 v[49:50], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; GFX9-NEXT:    v_lshrrev_b64 v[50:51], 24, v[11:12]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v30
; GFX9-NEXT:    v_lshrrev_b64 v[51:52], 24, v[9:10]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[3:4]
; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 24, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; GFX9-NEXT:    v_lshrrev_b64 v[52:53], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b64 v[42:43], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v29
; GFX9-NEXT:    v_mov_b32_e32 v43, v54
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[31:32]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v28
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v28
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[29:30]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v27
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v26
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[27:28]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v26
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v24
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[23:24]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v24
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v23
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v22
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v22
; GFX9-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v22
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[19:20]
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v3
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v1
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 16, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 24, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 8, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v17
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; GFX9-NEXT:  .LBB98_2: ; %Flow
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB98_4
; GFX9-NEXT:  ; %bb.3: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(28)
; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX9-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[29:30]
; GFX9-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[27:28]
; GFX9-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[25:26]
; GFX9-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[23:24]
; GFX9-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[33:34], 24, v[21:22]
; GFX9-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v16
; GFX9-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v16
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v15
; GFX9-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v15
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v14
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v32
; GFX9-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v14
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v32
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v31
; GFX9-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v13
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v31
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v30
; GFX9-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v12
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v30
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; GFX9-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v11
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v29
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v28
; GFX9-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v10
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v28
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; GFX9-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v9
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v27
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v26
; GFX9-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v8
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v26
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; GFX9-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v7
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v25
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v24
; GFX9-NEXT:    v_lshrrev_b64 v[48:49], 24, v[15:16]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v24
; GFX9-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[49:50], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v6
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v24
; GFX9-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[50:51], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v5
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; GFX9-NEXT:    v_lshrrev_b64 v[51:52], 24, v[9:10]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v23
; GFX9-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[52:53], 24, v[7:8]
; GFX9-NEXT:    v_lshrrev_b64 v[40:41], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v4
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v22
; GFX9-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[41:42], 24, v[3:4]
; GFX9-NEXT:    v_lshrrev_b64 v[53:54], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 24, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 16, v22
; GFX9-NEXT:    v_lshrrev_b64 v[42:43], 24, v[1:2]
; GFX9-NEXT:    v_lshrrev_b64 v[54:55], 24, v[17:18]
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 8, v22
; GFX9-NEXT:    v_lshrrev_b32_e32 v43, 24, v16
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 8, v5
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 8, v4
; GFX9-NEXT:    v_lshrrev_b32_e32 v62, 16, v3
; GFX9-NEXT:    v_lshrrev_b32_e32 v44, 8, v3
; GFX9-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v63, 16, v1
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 8, v1
; GFX9-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 16, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 8, v21
; GFX9-NEXT:    v_lshrrev_b32_e32 v34, 24, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v35, 16, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v36, 24, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v38, 16, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 8, v18
; GFX9-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v17
; GFX9-NEXT:  .LBB98_4: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v1, v1, v33, s4
; GFX9-NEXT:    v_perm_b32 v33, v62, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v55, v63, v42, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v55, 16, v1
; GFX9-NEXT:    v_perm_b32 v39, v39, v53, s4
; GFX9-NEXT:    v_perm_b32 v2, v2, v45, s4
; GFX9-NEXT:    v_perm_b32 v3, v3, v44, s4
; GFX9-NEXT:    v_perm_b32 v4, v4, v61, s4
; GFX9-NEXT:    v_perm_b32 v5, v5, v60, s4
; GFX9-NEXT:    v_perm_b32 v37, v37, v54, s4
; GFX9-NEXT:    v_perm_b32 v17, v17, v59, s4
; GFX9-NEXT:    v_perm_b32 v18, v18, v58, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v40, v41, v40, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v52, v41, v52, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v7, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v51, v41, v51, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v9, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v10, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v50, v41, v50, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v11, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v12, v12, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v49, v41, v49, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v14, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v48, v41, v48, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v15, v15, v41, s4
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v41, s4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v53, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshl_or_b32 v1, v33, 16, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    v_lshl_or_b32 v1, v40, 16, v5
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v6
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshl_or_b32 v1, v52, 16, v7
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v8
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshl_or_b32 v1, v51, 16, v9
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v10
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshl_or_b32 v1, v50, 16, v11
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v12
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshl_or_b32 v1, v49, 16, v13
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v14
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshl_or_b32 v1, v48, 16, v15
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v2, v35, v34, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v1, v43, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v16
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshl_or_b32 v1, v37, 16, v17
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    v_perm_b32 v1, v38, v36, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v1, 16, v18
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    v_perm_b32 v1, v19, v57, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    v_perm_b32 v1, v20, v56, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v1, v21, v47, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v2, v46, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v22, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v23, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v24, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v25, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v26, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v27, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v28, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v29, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v30, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v1, v31, v1, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v1, v32, v1, s4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64i16_to_v128i8:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x2
; GFX11-TRUE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr166_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr70_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr165_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr164_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr163_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr69_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr162_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr160_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr148_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr68_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr146_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr144_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr135_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr64_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr134_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr132_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr128_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr54_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr118_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr117_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr116_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr51_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr113_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr102_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr101_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr97_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr96_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr87_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr37_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr85_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr81_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr36_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr66_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr161_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr151_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr150_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr65_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr149_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr147_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr145_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr52_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr133_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr131_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr130_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr49_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr129_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr119_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr115_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr38_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr114_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr112_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr103_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr35_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr100_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr99_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr98_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr34_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr86_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr84_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr83_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr82_lo16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr80_lo16
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr33_lo16
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB98_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 8, v17
; GFX11-TRUE16-NEXT:  .LBB98_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB98_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[7:8]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[68:69], 24, v[5:6]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[31:32]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[15:16]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[13:14]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[51:52], 24, v[11:12]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[69:70], 24, v[3:4]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[29:30]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[19:20]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[9:10]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[70:71], 24, v[1:2]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[27:28]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[25:26]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[23:24]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[21:22]
; GFX11-TRUE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[17:18]
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 24, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 8, v16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 8, v15
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 24, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v101, 8, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v102, 24, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v113, 8, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v116, 8, v11
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v117, 24, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v118, 8, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v132, 24, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v134, 8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v135, 8, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v144, 24, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v148, 8, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v163, 8, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v164, 24, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v165, 8, v2
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 24, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 8, v32
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 8, v31
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v84, 24, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v30
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v98, 8, v29
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v99, 24, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v100, 8, v28
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v103, 8, v27
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v112, 24, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v114, 8, v26
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v25
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v119, 24, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v129, 8, v24
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v23
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v22
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v21
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v149, 8, v20
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v19
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v151, 24, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v161, 8, v18
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 8, v17
; GFX11-TRUE16-NEXT:  .LBB98_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v1.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v1, v1, v166, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v2.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v3.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v4.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v70, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v2, v2, v165, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v50, v50, v164, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v3, v3, v163, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v53, v69, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v5.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v162, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v160, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v50, 16, v2
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v53, 16, v3
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v5, v148, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v6.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v68, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v7.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v55, 16, v4
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v55.l, v8.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v146, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v50, v50, v144, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v135, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v53, v53, v64, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v39, 16, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v9.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v134, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v55, v55, v132, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v50, 16, v6
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v53, 16, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v50.l, v10.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v128, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v53.l, v11.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v39, v39, v54, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v55, 16, v8
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v118, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v50, v50, v117, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v116, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v51, v53, v51, 0xc0c0004
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v12.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v13.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v14.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v15.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v2, v50, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v3, v51, 16, v11
; GFX11-TRUE16-NEXT:    v_perm_b32 v5, v12, v113, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v4, v4, v102, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v13, v101, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v6, v6, v48, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v14, v97, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v7, v7, v96, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v15, v87, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v9, v37, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v5
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v8
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v16.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v6, v7, 16, v10
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v7, v9, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v17.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v18.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v19.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v9, v16, v85, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v8, v8, v81, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v17, v36, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v10, v10, v66, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v18, v161, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v11, v11, v151, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v19, v150, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v13, v65, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v8, v8, 16, v9
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v9, v10, 16, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v20.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v14
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v21.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v22.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v23.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v13, v20, v149, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v12, v12, v147, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v21, v145, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v14, v14, v52, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v22, v133, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v15, v15, v131, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v23, v130, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v17, v49, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v12, v12, 16, v13
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v13, v14, 16, v16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v24.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v14, v15, 16, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v25.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v15, v17, 16, v19
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v26.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v27.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v17, v24, v129, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v16, v16, v119, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v25, v115, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v18, v18, v38, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v26, v114, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v19, v19, v112, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v27, v103, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v35, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v17
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v17, v18, 16, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v28.h
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v18, v19, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v19, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v29.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v30.h
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v31.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v32.h
; GFX11-TRUE16-NEXT:    v_perm_b32 v22, v28, v100, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v20, v20, v99, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v23, v29, v98, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v21, v21, v34, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v26, v30, v86, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v24, v24, v84, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v28, v31, v83, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v25, v25, v33, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v29, v32, v82, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_perm_b32 v27, v27, v80, 0xc0c0004
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v20, v20, 16, v22
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v23
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v22, v24, 16, v26
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v23, v25, 16, v28
; GFX11-TRUE16-NEXT:    v_lshl_or_b32 v24, v27, 16, v29
; GFX11-TRUE16-NEXT:    s_clause 0x5
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-TRUE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x2
; GFX11-FAKE16-NEXT:    scratch_load_b32 v33, off, s32 offset:8
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_clause 0x14 ; 84-byte Folded Spill
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 offset:92
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v41, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v42, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v43, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v44, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v45, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v46, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v47, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v56, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v57, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v58, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v59, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v60, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v61, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v62, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v63, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v72, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v73, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v74, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v75, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v76, s32 offset:12
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr162
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr161
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr64
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr151
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr149
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr147
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr145
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr134
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr52
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr133
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr132
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr131
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr130
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr129
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr128
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr119
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr118
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr117
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr116
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr37
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr115
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr114
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr113
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr112
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr103
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr36
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr102
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr101
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr100
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr99
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr98
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr35
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr97
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr96
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr87
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr86
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr85
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr34
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr84
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr83
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr82
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr81
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr80
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr71
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr70
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr69
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr76
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr75
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr67
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr74
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr73
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr72
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr63
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr62
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr66
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr61
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr60
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr59
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr58
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr57
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr65
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr56
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr47
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr46
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr45
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr44
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr54
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr43
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr42
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr41
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr40
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr183
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr53
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr182
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr181
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr180
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr179
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr178
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr50
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr177
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr176
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr167
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr166
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr165
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr164
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr163
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr160
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr150
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr148
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr146
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr144
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr135
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr49
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr38
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(2)
; GFX11-FAKE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v33
; GFX11-FAKE16-NEXT:    ; implicit-def: $vgpr33
; GFX11-FAKE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB98_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v1
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v32
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[17:18]
; GFX11-FAKE16-NEXT:  .LBB98_2: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB98_4
; GFX11-FAKE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[33:34], 24, v[15:16]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[34:35], 24, v[13:14]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[35:36], 24, v[11:12]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[64:65], 24, v[1:2]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[36:37], 24, v[9:10]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[48:49], 24, v[5:6]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[52:53], 24, v[3:4]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[65:66], 24, v[21:22]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[37:38], 24, v[7:8]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[49:50], 24, v[29:30]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[53:54], 24, v[25:26]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[66:67], 24, v[19:20]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[38:39], 24, v[31:32]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[50:51], 24, v[27:28]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-FAKE16-NEXT:    v_lshrrev_b64 v[67:68], 24, v[17:18]
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 24, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 8, v16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 8, v15
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 24, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 8, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 8, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 24, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 8, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v99, 8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v100, 24, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v101, 16, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v102, 8, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v103, 16, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v112, 8, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v113, 24, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v114, 16, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v115, 8, v8
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v116, 16, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v117, 8, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v118, 24, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v119, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v128, 8, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v129, 16, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v130, 8, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v131, 24, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v132, 16, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v133, 8, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v134, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v145, 8, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v147, 24, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v149, 16, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v151, 8, v2
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v161, 16, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v162, 8, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v135, 24, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v144, 16, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v146, 8, v32
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v148, 16, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v150, 8, v31
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v160, 24, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v163, 16, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v164, 8, v30
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v165, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v166, 8, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v167, 24, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v176, 16, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v177, 8, v28
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v178, 16, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v179, 8, v27
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v180, 24, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v181, 16, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v182, 8, v26
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v183, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v40, 8, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v41, 24, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v42, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v43, 8, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v45, 8, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v46, 24, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v47, 16, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v56, 8, v22
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v57, 16, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v58, 8, v21
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v59, 24, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v60, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v61, 8, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v62, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v63, 8, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v72, 24, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v73, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v74, 8, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v75, 16, v17
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v76, 8, v17
; GFX11-FAKE16-NEXT:  .LBB98_4: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v161, v64, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v162, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v151, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v51, v149, v147, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v145, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v134, v52, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v39, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v133, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v132, v131, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v117, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v116, v37, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v51, 16, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v52, 16, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v130, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v48, v129, v48, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v128, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v51, v119, v118, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v115, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v52, v114, v113, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v4
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v37, 16, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v112, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v36, v103, v36, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v102, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v37, v101, v100, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v99, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v35, v98, v35, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v97, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v39, v96, v87, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v48, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v51, 16, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v52, 16, v8
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v36, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v37, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v35, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v39, 16, v12
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v13, v86, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v85, v34, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v14, v84, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v83, v82, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v15, v81, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v80, v33, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v16, v71, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v70, v69, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v17, v76, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v75, v67, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v6, 16, v5
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v8, 16, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v10, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v12, 16, v11
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v14, 16, v13
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v18, v74, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v73, v72, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v19, v63, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v62, v66, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v20, v61, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v60, v59, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v21, v58, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v57, v65, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v22, v56, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v47, v46, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v11, 16, v10
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v13, 16, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v15, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v17, 16, v16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v19, 16, v18
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v23, v45, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v44, v54, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v24, v43, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v42, v41, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v25, v40, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v183, v53, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v26, v182, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v181, v180, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v27, v179, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v178, v50, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v16, 16, v15
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v18, 16, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v20, 16, v19
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v22, 16, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v28, v177, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v176, v167, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v24, 16, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v29, v166, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v165, v49, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v30, v164, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v163, v160, 0xc0c0004
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v31, v150, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v148, v38, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v32, v146, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v144, v135, 0xc0c0004
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v21, 16, v20
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v23, 16, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v25, 16, v24
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v27, 16, v26
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v29, 16, v28
; GFX11-FAKE16-NEXT:    s_clause 0x5
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[1:4], off offset:32
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[5:8], off offset:48
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[9:12], off offset:64
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[13:16], off offset:80
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[17:20], off offset:96
; GFX11-FAKE16-NEXT:    scratch_store_b128 v0, v[21:24], off offset:112
; GFX11-FAKE16-NEXT:    s_clause 0x14 ; 84-byte Folded Reload
; GFX11-FAKE16-NEXT:    scratch_load_b32 v76, off, s32 offset:12
; GFX11-FAKE16-NEXT:    scratch_load_b32 v75, off, s32 offset:16
; GFX11-FAKE16-NEXT:    scratch_load_b32 v74, off, s32 offset:20
; GFX11-FAKE16-NEXT:    scratch_load_b32 v73, off, s32 offset:24
; GFX11-FAKE16-NEXT:    scratch_load_b32 v72, off, s32 offset:28
; GFX11-FAKE16-NEXT:    scratch_load_b32 v63, off, s32 offset:32
; GFX11-FAKE16-NEXT:    scratch_load_b32 v62, off, s32 offset:36
; GFX11-FAKE16-NEXT:    scratch_load_b32 v61, off, s32 offset:40
; GFX11-FAKE16-NEXT:    scratch_load_b32 v60, off, s32 offset:44
; GFX11-FAKE16-NEXT:    scratch_load_b32 v59, off, s32 offset:48
; GFX11-FAKE16-NEXT:    scratch_load_b32 v58, off, s32 offset:52
; GFX11-FAKE16-NEXT:    scratch_load_b32 v57, off, s32 offset:56
; GFX11-FAKE16-NEXT:    scratch_load_b32 v56, off, s32 offset:60
; GFX11-FAKE16-NEXT:    scratch_load_b32 v47, off, s32 offset:64
; GFX11-FAKE16-NEXT:    scratch_load_b32 v46, off, s32 offset:68
; GFX11-FAKE16-NEXT:    scratch_load_b32 v45, off, s32 offset:72
; GFX11-FAKE16-NEXT:    scratch_load_b32 v44, off, s32 offset:76
; GFX11-FAKE16-NEXT:    scratch_load_b32 v43, off, s32 offset:80
; GFX11-FAKE16-NEXT:    scratch_load_b32 v42, off, s32 offset:84
; GFX11-FAKE16-NEXT:    scratch_load_b32 v41, off, s32 offset:88
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 offset:92
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define inreg <128 x i8> @bitcast_v64i16_to_v128i8_scalar(<64 x i16> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64i16_to_v128i8_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_writelane_b32 v20, s30, 0
; SI-NEXT:    v_writelane_b32 v20, s31, 1
; SI-NEXT:    v_writelane_b32 v20, s34, 2
; SI-NEXT:    v_writelane_b32 v20, s35, 3
; SI-NEXT:    v_writelane_b32 v20, s36, 4
; SI-NEXT:    v_writelane_b32 v20, s37, 5
; SI-NEXT:    v_writelane_b32 v20, s38, 6
; SI-NEXT:    v_writelane_b32 v20, s39, 7
; SI-NEXT:    v_writelane_b32 v20, s48, 8
; SI-NEXT:    v_writelane_b32 v20, s49, 9
; SI-NEXT:    v_writelane_b32 v20, s50, 10
; SI-NEXT:    v_writelane_b32 v20, s51, 11
; SI-NEXT:    v_writelane_b32 v20, s52, 12
; SI-NEXT:    v_writelane_b32 v20, s53, 13
; SI-NEXT:    v_writelane_b32 v20, s54, 14
; SI-NEXT:    v_writelane_b32 v20, s55, 15
; SI-NEXT:    v_writelane_b32 v20, s64, 16
; SI-NEXT:    v_writelane_b32 v20, s65, 17
; SI-NEXT:    v_writelane_b32 v20, s66, 18
; SI-NEXT:    v_writelane_b32 v20, s67, 19
; SI-NEXT:    v_writelane_b32 v20, s68, 20
; SI-NEXT:    v_writelane_b32 v20, s69, 21
; SI-NEXT:    v_writelane_b32 v20, s70, 22
; SI-NEXT:    v_writelane_b32 v20, s71, 23
; SI-NEXT:    v_writelane_b32 v20, s80, 24
; SI-NEXT:    v_writelane_b32 v20, s81, 25
; SI-NEXT:    v_writelane_b32 v20, s82, 26
; SI-NEXT:    v_writelane_b32 v20, s83, 27
; SI-NEXT:    v_readfirstlane_b32 s37, v1
; SI-NEXT:    v_writelane_b32 v20, s84, 28
; SI-NEXT:    v_readfirstlane_b32 s39, v3
; SI-NEXT:    s_lshr_b32 s6, s37, 16
; SI-NEXT:    ; implicit-def: $vgpr21 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v20, s85, 29
; SI-NEXT:    v_readfirstlane_b32 s49, v5
; SI-NEXT:    s_lshr_b32 s7, s39, 16
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v21, s6, 0
; SI-NEXT:    v_writelane_b32 v20, s86, 30
; SI-NEXT:    v_readfirstlane_b32 s51, v7
; SI-NEXT:    s_lshr_b32 s8, s49, 16
; SI-NEXT:    v_writelane_b32 v21, s7, 1
; SI-NEXT:    v_writelane_b32 v20, s87, 31
; SI-NEXT:    v_readfirstlane_b32 s53, v9
; SI-NEXT:    s_lshr_b32 s9, s51, 16
; SI-NEXT:    v_writelane_b32 v21, s8, 2
; SI-NEXT:    v_writelane_b32 v20, s96, 32
; SI-NEXT:    v_readfirstlane_b32 s55, v11
; SI-NEXT:    s_lshr_b32 s10, s53, 16
; SI-NEXT:    v_writelane_b32 v21, s9, 3
; SI-NEXT:    v_writelane_b32 v20, s97, 33
; SI-NEXT:    v_readfirstlane_b32 s65, v13
; SI-NEXT:    s_lshr_b32 s11, s55, 16
; SI-NEXT:    v_writelane_b32 v21, s10, 4
; SI-NEXT:    v_writelane_b32 v20, s98, 34
; SI-NEXT:    v_readfirstlane_b32 s68, v18
; SI-NEXT:    v_readfirstlane_b32 s69, v17
; SI-NEXT:    v_readfirstlane_b32 s66, v16
; SI-NEXT:    v_readfirstlane_b32 s67, v15
; SI-NEXT:    v_readfirstlane_b32 s64, v14
; SI-NEXT:    s_lshr_b32 vcc_lo, s65, 16
; SI-NEXT:    v_readfirstlane_b32 s54, v12
; SI-NEXT:    v_readfirstlane_b32 s52, v10
; SI-NEXT:    v_readfirstlane_b32 s50, v8
; SI-NEXT:    v_readfirstlane_b32 s48, v6
; SI-NEXT:    v_readfirstlane_b32 s38, v4
; SI-NEXT:    v_readfirstlane_b32 s36, v2
; SI-NEXT:    v_writelane_b32 v21, s11, 5
; SI-NEXT:    v_writelane_b32 v20, s99, 35
; SI-NEXT:    s_lshr_b32 s92, s29, 16
; SI-NEXT:    s_lshr_b32 s99, s28, 16
; SI-NEXT:    s_lshr_b32 s93, s27, 16
; SI-NEXT:    s_lshr_b32 s97, s26, 16
; SI-NEXT:    s_lshr_b32 s94, s25, 16
; SI-NEXT:    s_lshr_b32 s87, s24, 16
; SI-NEXT:    s_lshr_b32 s30, s23, 16
; SI-NEXT:    s_lshr_b32 s85, s22, 16
; SI-NEXT:    s_lshr_b32 s31, s21, 16
; SI-NEXT:    s_lshr_b32 s83, s20, 16
; SI-NEXT:    s_lshr_b32 s34, s19, 16
; SI-NEXT:    s_lshr_b32 s81, s18, 16
; SI-NEXT:    s_lshr_b32 s35, s17, 16
; SI-NEXT:    s_lshr_b32 s71, s16, 16
; SI-NEXT:    s_lshr_b32 s76, s68, 16
; SI-NEXT:    s_lshr_b32 s70, s69, 16
; SI-NEXT:    s_lshr_b32 s77, s66, 16
; SI-NEXT:    s_lshr_b32 vcc_hi, s67, 16
; SI-NEXT:    s_lshr_b32 s78, s64, 16
; SI-NEXT:    s_lshr_b32 s79, s54, 16
; SI-NEXT:    s_lshr_b32 s88, s52, 16
; SI-NEXT:    s_lshr_b32 s89, s50, 16
; SI-NEXT:    s_lshr_b32 s90, s48, 16
; SI-NEXT:    s_lshr_b32 s91, s38, 16
; SI-NEXT:    s_lshr_b32 s95, s36, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v19
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 6
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 7
; SI-NEXT:    v_writelane_b32 v21, s70, 8
; SI-NEXT:    ; implicit-def: $vgpr22 : SGPR spill to VGPR lane
; SI-NEXT:    s_cbranch_scc0 .LBB99_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_and_b32 s4, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s71, 16
; SI-NEXT:    s_or_b32 s72, s4, s5
; SI-NEXT:    s_and_b32 s4, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s35, 16
; SI-NEXT:    s_or_b32 s73, s4, s5
; SI-NEXT:    s_and_b32 s4, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s81, 16
; SI-NEXT:    s_or_b32 s62, s4, s5
; SI-NEXT:    s_and_b32 s4, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s34, 16
; SI-NEXT:    s_or_b32 s63, s4, s5
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s83, 16
; SI-NEXT:    s_or_b32 s58, s4, s5
; SI-NEXT:    s_and_b32 s4, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s31, 16
; SI-NEXT:    s_or_b32 s59, s4, s5
; SI-NEXT:    s_and_b32 s4, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s85, 16
; SI-NEXT:    s_or_b32 s46, s4, s5
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s30, 16
; SI-NEXT:    s_or_b32 s47, s4, s5
; SI-NEXT:    s_and_b32 s4, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s87, 16
; SI-NEXT:    s_or_b32 s44, s4, s5
; SI-NEXT:    s_and_b32 s4, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s94, 16
; SI-NEXT:    s_or_b32 s45, s4, s5
; SI-NEXT:    s_and_b32 s4, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s97, 16
; SI-NEXT:    s_or_b32 s74, s4, s5
; SI-NEXT:    s_and_b32 s4, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s93, 16
; SI-NEXT:    s_or_b32 s75, s4, s5
; SI-NEXT:    s_and_b32 s4, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s99, 16
; SI-NEXT:    s_or_b32 s60, s4, s5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s92, 16
; SI-NEXT:    s_or_b32 s61, s4, s5
; SI-NEXT:    s_and_b32 s4, s37, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s6, 16
; SI-NEXT:    s_or_b32 s56, s4, s5
; SI-NEXT:    s_and_b32 s4, s36, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s95, 16
; SI-NEXT:    s_or_b32 s57, s4, s5
; SI-NEXT:    s_and_b32 s4, s39, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s7, 16
; SI-NEXT:    s_or_b32 s42, s4, s5
; SI-NEXT:    s_and_b32 s4, s38, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s91, 16
; SI-NEXT:    s_or_b32 s43, s4, s5
; SI-NEXT:    s_and_b32 s4, s49, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s8, 16
; SI-NEXT:    s_or_b32 s40, s4, s5
; SI-NEXT:    s_and_b32 s4, s48, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s90, 16
; SI-NEXT:    s_or_b32 s41, s4, s5
; SI-NEXT:    s_and_b32 s4, s51, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s9, 16
; SI-NEXT:    s_or_b32 s14, s4, s5
; SI-NEXT:    s_and_b32 s4, s50, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s89, 16
; SI-NEXT:    s_or_b32 s15, s4, s5
; SI-NEXT:    s_and_b32 s4, s53, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s10, 16
; SI-NEXT:    s_or_b32 s12, s4, s5
; SI-NEXT:    s_and_b32 s4, s52, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_or_b32 s13, s4, s5
; SI-NEXT:    s_and_b32 s4, s55, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s11, 16
; SI-NEXT:    s_or_b32 s10, s4, s5
; SI-NEXT:    s_and_b32 s4, s54, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s79, 16
; SI-NEXT:    s_or_b32 s11, s4, s5
; SI-NEXT:    s_and_b32 s4, s65, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_lo, 16
; SI-NEXT:    s_or_b32 s8, s4, s5
; SI-NEXT:    s_and_b32 s4, s64, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s78, 16
; SI-NEXT:    s_or_b32 s9, s4, s5
; SI-NEXT:    s_and_b32 s4, s67, 0xffff
; SI-NEXT:    s_lshl_b32 s5, vcc_hi, 16
; SI-NEXT:    s_or_b32 s6, s4, s5
; SI-NEXT:    s_and_b32 s4, s66, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s77, 16
; SI-NEXT:    s_or_b32 s7, s4, s5
; SI-NEXT:    s_and_b32 s4, s69, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s70, 16
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s68, 0xffff
; SI-NEXT:    s_lshl_b32 vcc_lo, s76, 16
; SI-NEXT:    s_or_b32 s5, s5, vcc_lo
; SI-NEXT:    s_lshr_b32 vcc_lo, s73, 8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 28
; SI-NEXT:    s_lshr_b32 vcc_lo, s63, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 30
; SI-NEXT:    s_lshr_b32 vcc_lo, s59, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 32
; SI-NEXT:    s_lshr_b32 vcc_lo, s47, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 34
; SI-NEXT:    s_lshr_b32 vcc_lo, s45, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 36
; SI-NEXT:    s_lshr_b32 vcc_lo, s75, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 38
; SI-NEXT:    s_lshr_b32 vcc_lo, s61, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 40
; SI-NEXT:    s_lshr_b32 vcc_lo, s57, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 42
; SI-NEXT:    s_lshr_b32 vcc_lo, s43, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 44
; SI-NEXT:    s_lshr_b32 vcc_lo, s41, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 46
; SI-NEXT:    s_lshr_b32 vcc_lo, s15, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 48
; SI-NEXT:    s_lshr_b32 vcc_lo, s13, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 50
; SI-NEXT:    s_lshr_b32 vcc_lo, s11, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 52
; SI-NEXT:    s_lshr_b32 vcc_lo, s9, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 54
; SI-NEXT:    s_lshr_b32 vcc_lo, s7, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 56
; SI-NEXT:    s_lshr_b32 vcc_lo, s5, 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 58
; SI-NEXT:    s_bfe_u32 vcc_lo, s35, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 27
; SI-NEXT:    s_bfe_u32 vcc_lo, s34, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 29
; SI-NEXT:    s_bfe_u32 vcc_lo, s31, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 31
; SI-NEXT:    s_bfe_u32 vcc_lo, s30, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 33
; SI-NEXT:    s_bfe_u32 vcc_lo, s94, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 35
; SI-NEXT:    s_bfe_u32 vcc_lo, s93, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 37
; SI-NEXT:    s_bfe_u32 vcc_lo, s92, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 39
; SI-NEXT:    s_bfe_u32 vcc_lo, s95, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 41
; SI-NEXT:    s_bfe_u32 vcc_lo, s91, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 43
; SI-NEXT:    s_bfe_u32 vcc_lo, s90, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 45
; SI-NEXT:    s_bfe_u32 vcc_lo, s89, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 47
; SI-NEXT:    s_bfe_u32 vcc_lo, s88, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 49
; SI-NEXT:    s_bfe_u32 vcc_lo, s79, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 51
; SI-NEXT:    s_bfe_u32 vcc_lo, s78, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 53
; SI-NEXT:    s_bfe_u32 vcc_lo, s77, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 55
; SI-NEXT:    s_bfe_u32 vcc_lo, s76, 0x80008
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 57
; SI-NEXT:    s_lshr_b64 vcc, s[72:73], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 13
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 14
; SI-NEXT:    s_lshr_b64 vcc, s[72:73], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 11
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 12
; SI-NEXT:    s_lshr_b64 vcc, s[72:73], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 9
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 10
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 19
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 20
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 17
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 18
; SI-NEXT:    s_lshr_b64 vcc, s[62:63], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 15
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 16
; SI-NEXT:    s_lshr_b64 vcc, s[58:59], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 25
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 26
; SI-NEXT:    s_lshr_b64 vcc, s[58:59], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 23
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 24
; SI-NEXT:    s_lshr_b64 vcc, s[58:59], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 21
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 22
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 31
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 32
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 29
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 30
; SI-NEXT:    s_lshr_b64 vcc, s[46:47], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 27
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 28
; SI-NEXT:    s_mov_b32 vcc_lo, s87
; SI-NEXT:    s_lshr_b64 s[86:87], s[44:45], 24
; SI-NEXT:    s_mov_b32 s87, vcc_lo
; SI-NEXT:    s_mov_b32 vcc_lo, s97
; SI-NEXT:    s_lshr_b64 s[96:97], s[44:45], 16
; SI-NEXT:    s_mov_b32 s97, vcc_lo
; SI-NEXT:    s_mov_b32 vcc_lo, s99
; SI-NEXT:    s_lshr_b64 s[98:99], s[44:45], 8
; SI-NEXT:    s_mov_b32 s99, vcc_lo
; SI-NEXT:    s_mov_b32 vcc_lo, s81
; SI-NEXT:    s_lshr_b64 s[80:81], s[74:75], 24
; SI-NEXT:    s_mov_b32 s81, vcc_lo
; SI-NEXT:    s_mov_b32 vcc_lo, s83
; SI-NEXT:    s_lshr_b64 s[82:83], s[74:75], 16
; SI-NEXT:    s_mov_b32 s83, vcc_lo
; SI-NEXT:    s_mov_b32 vcc_lo, s85
; SI-NEXT:    s_lshr_b64 s[84:85], s[74:75], 8
; SI-NEXT:    s_mov_b32 s85, vcc_lo
; SI-NEXT:    s_lshr_b64 vcc, s[60:61], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 37
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 38
; SI-NEXT:    s_lshr_b64 vcc, s[60:61], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 35
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 36
; SI-NEXT:    s_lshr_b64 vcc, s[60:61], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 33
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 34
; SI-NEXT:    s_lshr_b64 vcc, s[56:57], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 43
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 44
; SI-NEXT:    s_lshr_b64 vcc, s[56:57], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 41
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 42
; SI-NEXT:    s_lshr_b64 vcc, s[56:57], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 39
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 40
; SI-NEXT:    s_lshr_b64 vcc, s[42:43], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 49
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 50
; SI-NEXT:    s_lshr_b64 vcc, s[42:43], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 47
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 48
; SI-NEXT:    s_lshr_b64 vcc, s[42:43], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 45
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 46
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 55
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 56
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 53
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 54
; SI-NEXT:    s_lshr_b64 vcc, s[40:41], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 51
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 52
; SI-NEXT:    s_lshr_b64 vcc, s[14:15], 24
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 61
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 62
; SI-NEXT:    s_lshr_b64 vcc, s[14:15], 16
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 59
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 60
; SI-NEXT:    s_lshr_b64 vcc, s[14:15], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 57
; SI-NEXT:    v_writelane_b32 v21, vcc_hi, 58
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 24
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 3
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 4
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 16
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 1
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 2
; SI-NEXT:    s_lshr_b64 vcc, s[12:13], 8
; SI-NEXT:    v_writelane_b32 v21, vcc_lo, 63
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 0
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 24
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 9
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 10
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 16
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 7
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 8
; SI-NEXT:    s_lshr_b64 vcc, s[10:11], 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 5
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 6
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 24
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 15
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 16
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 16
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 13
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 14
; SI-NEXT:    s_lshr_b64 vcc, s[8:9], 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 11
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 12
; SI-NEXT:    s_lshr_b64 vcc, s[6:7], 24
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 21
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 22
; SI-NEXT:    s_lshr_b64 vcc, s[6:7], 16
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 19
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 20
; SI-NEXT:    s_lshr_b64 vcc, s[6:7], 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 17
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 18
; SI-NEXT:    s_lshr_b64 vcc, s[4:5], 24
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 25
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 26
; SI-NEXT:    s_mov_b32 vcc_lo, s71
; SI-NEXT:    s_lshr_b64 s[70:71], s[4:5], 16
; SI-NEXT:    s_mov_b32 s71, vcc_lo
; SI-NEXT:    s_lshr_b64 vcc, s[4:5], 8
; SI-NEXT:    v_writelane_b32 v22, vcc_lo, 23
; SI-NEXT:    v_writelane_b32 v22, vcc_hi, 24
; SI-NEXT:    s_cbranch_execnz .LBB99_3
; SI-NEXT:  .LBB99_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s69, s69, 3
; SI-NEXT:    v_readlane_b32 s5, v21, 8
; SI-NEXT:    s_and_b32 s4, s69, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_add_i32 s68, s68, 3
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s6, s76, 16
; SI-NEXT:    s_add_i32 s67, s67, 3
; SI-NEXT:    v_readlane_b32 s7, v21, 7
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    s_and_b32 s6, s67, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_add_i32 s66, s66, 3
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    s_and_b32 s7, s66, 0xffff
; SI-NEXT:    s_lshl_b32 s8, s77, 16
; SI-NEXT:    s_add_i32 s65, s65, 3
; SI-NEXT:    v_readlane_b32 s9, v21, 6
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    s_and_b32 s8, s65, 0xffff
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_add_i32 s64, s64, 3
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    s_and_b32 s9, s64, 0xffff
; SI-NEXT:    s_lshl_b32 s10, s78, 16
; SI-NEXT:    s_add_i32 s55, s55, 3
; SI-NEXT:    v_readlane_b32 s11, v21, 5
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    s_and_b32 s10, s55, 0xffff
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_add_i32 s54, s54, 3
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    s_and_b32 s11, s54, 0xffff
; SI-NEXT:    s_lshl_b32 s12, s79, 16
; SI-NEXT:    s_add_i32 s53, s53, 3
; SI-NEXT:    v_readlane_b32 s13, v21, 4
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    s_and_b32 s12, s53, 0xffff
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_add_i32 s52, s52, 3
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    s_and_b32 s13, s52, 0xffff
; SI-NEXT:    s_lshl_b32 s14, s88, 16
; SI-NEXT:    s_add_i32 s51, s51, 3
; SI-NEXT:    v_readlane_b32 s15, v21, 3
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    s_and_b32 s14, s51, 0xffff
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_add_i32 s50, s50, 3
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    s_and_b32 s15, s50, 0xffff
; SI-NEXT:    s_lshl_b32 s40, s89, 16
; SI-NEXT:    s_add_i32 s49, s49, 3
; SI-NEXT:    v_readlane_b32 s41, v21, 2
; SI-NEXT:    s_or_b32 s15, s40, s15
; SI-NEXT:    s_and_b32 s40, s49, 0xffff
; SI-NEXT:    s_lshl_b32 s41, s41, 16
; SI-NEXT:    s_add_i32 s48, s48, 3
; SI-NEXT:    s_or_b32 s40, s41, s40
; SI-NEXT:    s_and_b32 s41, s48, 0xffff
; SI-NEXT:    s_lshl_b32 s42, s90, 16
; SI-NEXT:    s_add_i32 s39, s39, 3
; SI-NEXT:    v_readlane_b32 s43, v21, 1
; SI-NEXT:    s_or_b32 s41, s42, s41
; SI-NEXT:    s_and_b32 s42, s39, 0xffff
; SI-NEXT:    s_lshl_b32 s43, s43, 16
; SI-NEXT:    s_add_i32 s38, s38, 3
; SI-NEXT:    s_or_b32 s42, s43, s42
; SI-NEXT:    s_and_b32 s43, s38, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s91, 16
; SI-NEXT:    s_add_i32 s37, s37, 3
; SI-NEXT:    v_readlane_b32 s45, v21, 0
; SI-NEXT:    s_or_b32 s43, s44, s43
; SI-NEXT:    s_and_b32 s44, s37, 0xffff
; SI-NEXT:    s_lshl_b32 s45, s45, 16
; SI-NEXT:    s_or_b32 s44, s45, s44
; SI-NEXT:    s_add_i32 s36, s36, 3
; SI-NEXT:    s_add_i32 s56, s44, 0x30000
; SI-NEXT:    s_and_b32 s44, s36, 0xffff
; SI-NEXT:    s_lshl_b32 s45, s95, 16
; SI-NEXT:    s_or_b32 s44, s45, s44
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    s_add_i32 s57, s44, 0x30000
; SI-NEXT:    s_and_b32 s28, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s99, 16
; SI-NEXT:    s_or_b32 s28, s44, s28
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    s_add_i32 s60, s28, 0x30000
; SI-NEXT:    s_and_b32 s28, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s29, s92, 16
; SI-NEXT:    s_or_b32 s28, s29, s28
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    s_add_i32 s61, s28, 0x30000
; SI-NEXT:    s_and_b32 s26, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s28, s97, 16
; SI-NEXT:    s_or_b32 s26, s28, s26
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    s_add_i32 s74, s26, 0x30000
; SI-NEXT:    s_and_b32 s26, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s27, s93, 16
; SI-NEXT:    s_or_b32 s26, s27, s26
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    s_add_i32 s75, s26, 0x30000
; SI-NEXT:    s_and_b32 s24, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s26, s87, 16
; SI-NEXT:    s_or_b32 s24, s26, s24
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_add_i32 s44, s24, 0x30000
; SI-NEXT:    s_and_b32 s24, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s94, 16
; SI-NEXT:    s_or_b32 s24, s25, s24
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    s_add_i32 s45, s24, 0x30000
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s24, s85, 16
; SI-NEXT:    s_or_b32 s22, s24, s22
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    s_add_i32 s46, s22, 0x30000
; SI-NEXT:    s_and_b32 s22, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s23, s30, 16
; SI-NEXT:    s_or_b32 s22, s23, s22
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    s_add_i32 s47, s22, 0x30000
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s22, s83, 16
; SI-NEXT:    s_or_b32 s20, s22, s20
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    s_add_i32 s58, s20, 0x30000
; SI-NEXT:    s_and_b32 s20, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s21, s31, 16
; SI-NEXT:    s_or_b32 s20, s21, s20
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    s_add_i32 s59, s20, 0x30000
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s20, s81, 16
; SI-NEXT:    s_or_b32 s18, s20, s18
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    s_add_i32 s62, s18, 0x30000
; SI-NEXT:    s_and_b32 s18, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s19, s34, 16
; SI-NEXT:    s_or_b32 s18, s19, s18
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_add_i32 s63, s18, 0x30000
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s18, s71, 16
; SI-NEXT:    s_or_b32 s16, s18, s16
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_add_i32 s72, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s35, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s73, s16, 0x30000
; SI-NEXT:    s_lshr_b64 s[16:17], s[72:73], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 13
; SI-NEXT:    v_writelane_b32 v21, s17, 14
; SI-NEXT:    s_lshr_b64 s[16:17], s[72:73], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 11
; SI-NEXT:    v_writelane_b32 v21, s17, 12
; SI-NEXT:    s_lshr_b32 s16, s73, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 27
; SI-NEXT:    s_lshr_b32 s16, s73, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 28
; SI-NEXT:    s_lshr_b32 s16, s63, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 29
; SI-NEXT:    s_lshr_b32 s16, s63, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 30
; SI-NEXT:    s_lshr_b32 s16, s59, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 31
; SI-NEXT:    s_lshr_b32 s16, s59, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 32
; SI-NEXT:    s_lshr_b32 s16, s47, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 33
; SI-NEXT:    s_lshr_b32 s16, s47, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 34
; SI-NEXT:    s_lshr_b32 s16, s45, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 35
; SI-NEXT:    s_lshr_b32 s16, s45, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 36
; SI-NEXT:    s_lshr_b32 s16, s75, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 37
; SI-NEXT:    s_lshr_b32 s16, s75, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 38
; SI-NEXT:    s_lshr_b32 s16, s61, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 39
; SI-NEXT:    s_lshr_b32 s16, s61, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 40
; SI-NEXT:    s_lshr_b32 s16, s57, 24
; SI-NEXT:    s_add_i32 s43, s43, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 41
; SI-NEXT:    s_lshr_b32 s16, s57, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 42
; SI-NEXT:    s_lshr_b32 s16, s43, 24
; SI-NEXT:    s_add_i32 s41, s41, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 43
; SI-NEXT:    s_lshr_b32 s16, s43, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 44
; SI-NEXT:    s_lshr_b32 s16, s41, 24
; SI-NEXT:    s_add_i32 s15, s15, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 45
; SI-NEXT:    s_lshr_b32 s16, s41, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 46
; SI-NEXT:    s_lshr_b32 s16, s15, 24
; SI-NEXT:    s_add_i32 s13, s13, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 47
; SI-NEXT:    s_lshr_b32 s16, s15, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 48
; SI-NEXT:    s_lshr_b32 s16, s13, 24
; SI-NEXT:    s_add_i32 s11, s11, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 49
; SI-NEXT:    s_lshr_b32 s16, s13, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 50
; SI-NEXT:    s_lshr_b32 s16, s11, 24
; SI-NEXT:    s_add_i32 s9, s9, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 51
; SI-NEXT:    s_lshr_b32 s16, s11, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 52
; SI-NEXT:    s_lshr_b32 s16, s9, 24
; SI-NEXT:    s_add_i32 s7, s7, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 53
; SI-NEXT:    s_lshr_b32 s16, s9, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 54
; SI-NEXT:    s_lshr_b32 s16, s7, 24
; SI-NEXT:    s_add_i32 s5, s5, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 55
; SI-NEXT:    s_lshr_b32 s16, s7, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 56
; SI-NEXT:    s_lshr_b32 s16, s5, 24
; SI-NEXT:    v_writelane_b32 v22, s16, 57
; SI-NEXT:    s_lshr_b32 s16, s5, 8
; SI-NEXT:    v_writelane_b32 v22, s16, 58
; SI-NEXT:    s_lshr_b64 s[16:17], s[72:73], 8
; SI-NEXT:    v_writelane_b32 v21, s16, 9
; SI-NEXT:    v_writelane_b32 v21, s17, 10
; SI-NEXT:    s_lshr_b64 s[16:17], s[62:63], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 19
; SI-NEXT:    v_writelane_b32 v21, s17, 20
; SI-NEXT:    s_lshr_b64 s[16:17], s[62:63], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 17
; SI-NEXT:    v_writelane_b32 v21, s17, 18
; SI-NEXT:    s_lshr_b64 s[16:17], s[62:63], 8
; SI-NEXT:    v_writelane_b32 v21, s16, 15
; SI-NEXT:    v_writelane_b32 v21, s17, 16
; SI-NEXT:    s_lshr_b64 s[16:17], s[58:59], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 25
; SI-NEXT:    v_writelane_b32 v21, s17, 26
; SI-NEXT:    s_lshr_b64 s[16:17], s[58:59], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 23
; SI-NEXT:    v_writelane_b32 v21, s17, 24
; SI-NEXT:    s_lshr_b64 s[16:17], s[58:59], 8
; SI-NEXT:    v_writelane_b32 v21, s16, 21
; SI-NEXT:    v_writelane_b32 v21, s17, 22
; SI-NEXT:    s_lshr_b64 s[16:17], s[46:47], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 31
; SI-NEXT:    v_writelane_b32 v21, s17, 32
; SI-NEXT:    s_lshr_b64 s[16:17], s[46:47], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 29
; SI-NEXT:    v_writelane_b32 v21, s17, 30
; SI-NEXT:    s_lshr_b64 s[16:17], s[46:47], 8
; SI-NEXT:    v_writelane_b32 v21, s16, 27
; SI-NEXT:    v_writelane_b32 v21, s17, 28
; SI-NEXT:    s_lshr_b64 s[16:17], s[60:61], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 37
; SI-NEXT:    v_writelane_b32 v21, s17, 38
; SI-NEXT:    s_lshr_b64 s[16:17], s[60:61], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 35
; SI-NEXT:    v_writelane_b32 v21, s17, 36
; SI-NEXT:    s_lshr_b64 s[16:17], s[60:61], 8
; SI-NEXT:    v_writelane_b32 v21, s16, 33
; SI-NEXT:    v_writelane_b32 v21, s17, 34
; SI-NEXT:    s_lshr_b64 s[16:17], s[56:57], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 43
; SI-NEXT:    v_writelane_b32 v21, s17, 44
; SI-NEXT:    s_lshr_b64 s[16:17], s[56:57], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 41
; SI-NEXT:    v_writelane_b32 v21, s17, 42
; SI-NEXT:    s_lshr_b64 s[16:17], s[56:57], 8
; SI-NEXT:    s_add_i32 s42, s42, 0x30000
; SI-NEXT:    v_writelane_b32 v21, s16, 39
; SI-NEXT:    v_writelane_b32 v21, s17, 40
; SI-NEXT:    s_lshr_b64 s[16:17], s[42:43], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 49
; SI-NEXT:    v_writelane_b32 v21, s17, 50
; SI-NEXT:    s_lshr_b64 s[16:17], s[42:43], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 47
; SI-NEXT:    v_writelane_b32 v21, s17, 48
; SI-NEXT:    s_lshr_b64 s[16:17], s[42:43], 8
; SI-NEXT:    s_add_i32 s40, s40, 0x30000
; SI-NEXT:    v_writelane_b32 v21, s16, 45
; SI-NEXT:    v_writelane_b32 v21, s17, 46
; SI-NEXT:    s_lshr_b64 s[16:17], s[40:41], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 55
; SI-NEXT:    v_writelane_b32 v21, s17, 56
; SI-NEXT:    s_lshr_b64 s[16:17], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 53
; SI-NEXT:    v_writelane_b32 v21, s17, 54
; SI-NEXT:    s_lshr_b64 s[16:17], s[40:41], 8
; SI-NEXT:    s_add_i32 s14, s14, 0x30000
; SI-NEXT:    v_writelane_b32 v21, s16, 51
; SI-NEXT:    v_writelane_b32 v21, s17, 52
; SI-NEXT:    s_lshr_b64 s[16:17], s[14:15], 24
; SI-NEXT:    v_writelane_b32 v21, s16, 61
; SI-NEXT:    v_writelane_b32 v21, s17, 62
; SI-NEXT:    s_lshr_b64 s[16:17], s[14:15], 16
; SI-NEXT:    v_writelane_b32 v21, s16, 59
; SI-NEXT:    v_writelane_b32 v21, s17, 60
; SI-NEXT:    s_lshr_b64 s[16:17], s[14:15], 8
; SI-NEXT:    s_add_i32 s12, s12, 0x30000
; SI-NEXT:    v_writelane_b32 v21, s16, 57
; SI-NEXT:    v_writelane_b32 v21, s17, 58
; SI-NEXT:    s_lshr_b64 s[16:17], s[12:13], 24
; SI-NEXT:    v_writelane_b32 v22, s16, 3
; SI-NEXT:    v_writelane_b32 v22, s17, 4
; SI-NEXT:    s_lshr_b64 s[16:17], s[12:13], 16
; SI-NEXT:    v_writelane_b32 v22, s16, 1
; SI-NEXT:    v_writelane_b32 v22, s17, 2
; SI-NEXT:    s_lshr_b64 s[16:17], s[12:13], 8
; SI-NEXT:    s_add_i32 s10, s10, 0x30000
; SI-NEXT:    v_writelane_b32 v21, s16, 63
; SI-NEXT:    v_writelane_b32 v22, s17, 0
; SI-NEXT:    s_lshr_b64 s[16:17], s[10:11], 24
; SI-NEXT:    v_writelane_b32 v22, s16, 9
; SI-NEXT:    v_writelane_b32 v22, s17, 10
; SI-NEXT:    s_lshr_b64 s[16:17], s[10:11], 16
; SI-NEXT:    v_writelane_b32 v22, s16, 7
; SI-NEXT:    v_writelane_b32 v22, s17, 8
; SI-NEXT:    s_lshr_b64 s[16:17], s[10:11], 8
; SI-NEXT:    s_add_i32 s8, s8, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 5
; SI-NEXT:    v_writelane_b32 v22, s17, 6
; SI-NEXT:    s_lshr_b64 s[16:17], s[8:9], 24
; SI-NEXT:    v_writelane_b32 v22, s16, 15
; SI-NEXT:    v_writelane_b32 v22, s17, 16
; SI-NEXT:    s_lshr_b64 s[16:17], s[8:9], 16
; SI-NEXT:    v_writelane_b32 v22, s16, 13
; SI-NEXT:    v_writelane_b32 v22, s17, 14
; SI-NEXT:    s_lshr_b64 s[16:17], s[8:9], 8
; SI-NEXT:    s_add_i32 s6, s6, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 11
; SI-NEXT:    v_writelane_b32 v22, s17, 12
; SI-NEXT:    s_lshr_b64 s[16:17], s[6:7], 24
; SI-NEXT:    v_writelane_b32 v22, s16, 21
; SI-NEXT:    v_writelane_b32 v22, s17, 22
; SI-NEXT:    s_lshr_b64 s[16:17], s[6:7], 16
; SI-NEXT:    v_writelane_b32 v22, s16, 19
; SI-NEXT:    v_writelane_b32 v22, s17, 20
; SI-NEXT:    s_lshr_b64 s[16:17], s[6:7], 8
; SI-NEXT:    s_add_i32 s4, s4, 0x30000
; SI-NEXT:    v_writelane_b32 v22, s16, 17
; SI-NEXT:    v_writelane_b32 v22, s17, 18
; SI-NEXT:    s_lshr_b64 s[16:17], s[4:5], 24
; SI-NEXT:    v_writelane_b32 v22, s16, 25
; SI-NEXT:    v_writelane_b32 v22, s17, 26
; SI-NEXT:    s_lshr_b64 s[16:17], s[4:5], 8
; SI-NEXT:    s_lshr_b32 s35, s73, 16
; SI-NEXT:    s_lshr_b32 s34, s63, 16
; SI-NEXT:    s_lshr_b32 s31, s59, 16
; SI-NEXT:    s_lshr_b32 s30, s47, 16
; SI-NEXT:    s_lshr_b32 s94, s45, 16
; SI-NEXT:    s_lshr_b32 s93, s75, 16
; SI-NEXT:    s_lshr_b32 s92, s61, 16
; SI-NEXT:    s_lshr_b32 s95, s57, 16
; SI-NEXT:    s_lshr_b32 s91, s43, 16
; SI-NEXT:    s_lshr_b32 s90, s41, 16
; SI-NEXT:    s_lshr_b32 s89, s15, 16
; SI-NEXT:    s_lshr_b32 s88, s13, 16
; SI-NEXT:    s_lshr_b32 s79, s11, 16
; SI-NEXT:    s_lshr_b32 s78, s9, 16
; SI-NEXT:    s_lshr_b32 s77, s7, 16
; SI-NEXT:    s_lshr_b32 s76, s5, 16
; SI-NEXT:    s_lshr_b64 s[86:87], s[44:45], 24
; SI-NEXT:    s_lshr_b64 s[96:97], s[44:45], 16
; SI-NEXT:    s_lshr_b64 s[98:99], s[44:45], 8
; SI-NEXT:    s_lshr_b64 s[80:81], s[74:75], 24
; SI-NEXT:    s_lshr_b64 s[82:83], s[74:75], 16
; SI-NEXT:    s_lshr_b64 s[84:85], s[74:75], 8
; SI-NEXT:    s_lshr_b64 s[70:71], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v22, s16, 23
; SI-NEXT:    v_writelane_b32 v22, s17, 24
; SI-NEXT:  .LBB99_3: ; %end
; SI-NEXT:    v_readlane_b32 s16, v21, 9
; SI-NEXT:    v_readlane_b32 s17, v21, 10
; SI-NEXT:    v_readlane_b32 s18, v21, 11
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s72, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 12
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 13
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 28
; SI-NEXT:    v_mov_b32_e32 v1, s16
; SI-NEXT:    s_and_b32 s16, s73, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s35, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 27
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s19, v21, 14
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 15
; SI-NEXT:    v_readlane_b32 s17, v21, 16
; SI-NEXT:    v_readlane_b32 s18, v21, 17
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s62, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 18
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 19
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 30
; SI-NEXT:    v_mov_b32_e32 v3, s16
; SI-NEXT:    s_and_b32 s16, s63, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s34, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 29
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s19, v21, 20
; SI-NEXT:    v_mov_b32_e32 v4, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 21
; SI-NEXT:    v_readlane_b32 s17, v21, 22
; SI-NEXT:    v_readlane_b32 s18, v21, 23
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s58, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 24
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 25
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 32
; SI-NEXT:    v_mov_b32_e32 v5, s16
; SI-NEXT:    s_and_b32 s16, s59, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s31, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 31
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s19, v21, 26
; SI-NEXT:    v_mov_b32_e32 v6, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 27
; SI-NEXT:    v_readlane_b32 s17, v21, 28
; SI-NEXT:    v_readlane_b32 s18, v21, 29
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s46, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 30
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 31
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 34
; SI-NEXT:    v_mov_b32_e32 v7, s16
; SI-NEXT:    s_and_b32 s16, s47, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s30, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 33
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v8, s16
; SI-NEXT:    s_lshl_b32 s16, s98, 8
; SI-NEXT:    s_and_b32 s17, s44, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s96, 0xff
; SI-NEXT:    s_lshl_b32 s18, s86, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 36
; SI-NEXT:    v_mov_b32_e32 v9, s16
; SI-NEXT:    s_and_b32 s16, s45, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s94, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 35
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_mov_b32_e32 v10, s16
; SI-NEXT:    s_lshl_b32 s16, s84, 8
; SI-NEXT:    s_and_b32 s17, s74, 0xff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s82, 0xff
; SI-NEXT:    s_lshl_b32 s18, s80, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 38
; SI-NEXT:    v_mov_b32_e32 v11, s16
; SI-NEXT:    s_and_b32 s16, s75, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s93, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 37
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s19, v21, 32
; SI-NEXT:    v_mov_b32_e32 v12, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 33
; SI-NEXT:    v_readlane_b32 s17, v21, 34
; SI-NEXT:    v_readlane_b32 s18, v21, 35
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s60, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 36
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 37
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 40
; SI-NEXT:    v_mov_b32_e32 v13, s16
; SI-NEXT:    s_and_b32 s16, s61, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s92, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 39
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s19, v21, 38
; SI-NEXT:    v_mov_b32_e32 v14, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 39
; SI-NEXT:    v_readlane_b32 s17, v21, 40
; SI-NEXT:    v_readlane_b32 s18, v21, 41
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s56, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 42
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 43
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 42
; SI-NEXT:    v_mov_b32_e32 v15, s16
; SI-NEXT:    s_and_b32 s16, s57, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s95, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 41
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_i32_e32 v1, vcc, 4, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 8, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v3, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 12, v0
; SI-NEXT:    v_readlane_b32 s19, v21, 44
; SI-NEXT:    v_mov_b32_e32 v16, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 45
; SI-NEXT:    buffer_store_dword v4, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 16, v0
; SI-NEXT:    v_readlane_b32 s17, v21, 46
; SI-NEXT:    v_readlane_b32 s18, v21, 47
; SI-NEXT:    buffer_store_dword v5, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 20, v0
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    s_and_b32 s17, s42, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 48
; SI-NEXT:    buffer_store_dword v6, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 24, v0
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 49
; SI-NEXT:    buffer_store_dword v7, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 28, v0
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    buffer_store_dword v8, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 32, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    buffer_store_dword v9, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 36, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v10, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 40, v0
; SI-NEXT:    v_readlane_b32 s17, v22, 44
; SI-NEXT:    buffer_store_dword v11, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 44, v0
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s43, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    buffer_store_dword v12, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 48, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s91, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 43
; SI-NEXT:    v_readlane_b32 s19, v21, 50
; SI-NEXT:    buffer_store_dword v13, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 52, v0
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    buffer_store_dword v14, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 56, v0
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_readlane_b32 s18, v21, 51
; SI-NEXT:    buffer_store_dword v15, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 60, v0
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    v_readlane_b32 s19, v21, 52
; SI-NEXT:    buffer_store_dword v16, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v1, vcc, 64, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_lshl_b32 s17, s18, 8
; SI-NEXT:    v_readlane_b32 s18, v21, 53
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s40, 0xff
; SI-NEXT:    v_readlane_b32 s19, v21, 54
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s18, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 55
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x44, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    v_readlane_b32 s17, v22, 46
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    s_and_b32 s16, s41, 0xff
; SI-NEXT:    s_lshl_b32 s17, s17, 8
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s90, 0xff
; SI-NEXT:    v_readlane_b32 s18, v22, 45
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_lshl_b32 s18, s18, 24
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_or_b32 s17, s18, s17
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x48, v0
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 57
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s17, v21, 58
; SI-NEXT:    s_lshl_b32 s16, s16, 8
; SI-NEXT:    v_readlane_b32 s19, v21, 56
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    v_readlane_b32 s16, v21, 59
; SI-NEXT:    v_readlane_b32 s17, v21, 60
; SI-NEXT:    s_and_b32 s16, s16, 0xff
; SI-NEXT:    v_readlane_b32 s18, v21, 61
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s17, s18, 24
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x4c, v0
; SI-NEXT:    s_or_b32 s14, s14, s16
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s14
; SI-NEXT:    s_and_b32 s14, s15, 0xff
; SI-NEXT:    v_readlane_b32 s15, v22, 48
; SI-NEXT:    s_lshl_b32 s15, s15, 8
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_and_b32 s15, s89, 0xff
; SI-NEXT:    v_readlane_b32 s16, v22, 47
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_lshl_b32 s16, s16, 24
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x50, v0
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s14
; SI-NEXT:    v_readlane_b32 s14, v21, 63
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s15, v22, 0
; SI-NEXT:    s_lshl_b32 s14, s14, 8
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    v_readlane_b32 s14, v22, 1
; SI-NEXT:    v_readlane_b32 s15, v22, 2
; SI-NEXT:    s_and_b32 s14, s14, 0xff
; SI-NEXT:    v_readlane_b32 s16, v22, 3
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_lshl_b32 s15, s16, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x54, v0
; SI-NEXT:    s_or_b32 s12, s12, s14
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s12
; SI-NEXT:    s_and_b32 s12, s13, 0xff
; SI-NEXT:    v_readlane_b32 s13, v22, 50
; SI-NEXT:    s_lshl_b32 s13, s13, 8
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_and_b32 s13, s88, 0xff
; SI-NEXT:    v_readlane_b32 s14, v22, 49
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_lshl_b32 s14, s14, 24
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x58, v0
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s12
; SI-NEXT:    v_readlane_b32 s12, v22, 5
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s13, v22, 6
; SI-NEXT:    s_lshl_b32 s12, s12, 8
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    v_readlane_b32 s12, v22, 7
; SI-NEXT:    v_readlane_b32 s13, v22, 8
; SI-NEXT:    s_and_b32 s12, s12, 0xff
; SI-NEXT:    v_readlane_b32 s14, v22, 9
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_lshl_b32 s13, s14, 24
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x5c, v0
; SI-NEXT:    s_or_b32 s10, s10, s12
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s10
; SI-NEXT:    s_and_b32 s10, s11, 0xff
; SI-NEXT:    v_readlane_b32 s11, v22, 52
; SI-NEXT:    s_lshl_b32 s11, s11, 8
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_and_b32 s11, s79, 0xff
; SI-NEXT:    v_readlane_b32 s12, v22, 51
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_lshl_b32 s12, s12, 24
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x60, v0
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s10
; SI-NEXT:    v_readlane_b32 s10, v22, 11
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s11, v22, 12
; SI-NEXT:    s_lshl_b32 s10, s10, 8
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    v_readlane_b32 s10, v22, 13
; SI-NEXT:    v_readlane_b32 s11, v22, 14
; SI-NEXT:    s_and_b32 s10, s10, 0xff
; SI-NEXT:    v_readlane_b32 s12, v22, 15
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_lshl_b32 s11, s12, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x64, v0
; SI-NEXT:    s_or_b32 s8, s8, s10
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    s_and_b32 s8, s9, 0xff
; SI-NEXT:    v_readlane_b32 s9, v22, 54
; SI-NEXT:    s_lshl_b32 s9, s9, 8
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s9, s78, 0xff
; SI-NEXT:    v_readlane_b32 s10, v22, 53
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_lshl_b32 s10, s10, 24
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x68, v0
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s8
; SI-NEXT:    v_readlane_b32 s8, v22, 17
; SI-NEXT:    s_and_b32 s6, s6, 0xff
; SI-NEXT:    v_readlane_b32 s9, v22, 18
; SI-NEXT:    s_lshl_b32 s8, s8, 8
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    v_readlane_b32 s8, v22, 19
; SI-NEXT:    v_readlane_b32 s9, v22, 20
; SI-NEXT:    s_and_b32 s8, s8, 0xff
; SI-NEXT:    v_readlane_b32 s10, v22, 21
; SI-NEXT:    s_lshl_b32 s8, s8, 16
; SI-NEXT:    s_lshl_b32 s9, s10, 24
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x6c, v0
; SI-NEXT:    s_or_b32 s6, s6, s8
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    s_and_b32 s6, s7, 0xff
; SI-NEXT:    v_readlane_b32 s7, v22, 56
; SI-NEXT:    s_lshl_b32 s7, s7, 8
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_and_b32 s7, s77, 0xff
; SI-NEXT:    v_readlane_b32 s8, v22, 55
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_lshl_b32 s8, s8, 24
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x70, v0
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s6
; SI-NEXT:    v_readlane_b32 s6, v22, 23
; SI-NEXT:    s_and_b32 s4, s4, 0xff
; SI-NEXT:    s_lshl_b32 s6, s6, 8
; SI-NEXT:    v_readlane_b32 s7, v22, 24
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    s_and_b32 s6, s70, 0xff
; SI-NEXT:    v_readlane_b32 s8, v22, 25
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_lshl_b32 s7, s8, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x74, v0
; SI-NEXT:    s_or_b32 s4, s4, s6
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    s_and_b32 s4, s5, 0xff
; SI-NEXT:    v_readlane_b32 s5, v22, 58
; SI-NEXT:    s_lshl_b32 s5, s5, 8
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    s_and_b32 s5, s76, 0xff
; SI-NEXT:    v_readlane_b32 s6, v22, 57
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_lshl_b32 s6, s6, 24
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    v_add_i32_e32 v1, vcc, 0x78, v0
; SI-NEXT:    s_or_b32 s4, s4, s5
; SI-NEXT:    buffer_store_dword v2, v1, s[0:3], 0 offen
; SI-NEXT:    v_add_i32_e32 v0, vcc, 0x7c, v0
; SI-NEXT:    v_mov_b32_e32 v1, s4
; SI-NEXT:    v_readlane_b32 s19, v21, 62
; SI-NEXT:    v_readlane_b32 s17, v22, 4
; SI-NEXT:    v_readlane_b32 s15, v22, 10
; SI-NEXT:    v_readlane_b32 s13, v22, 16
; SI-NEXT:    v_readlane_b32 s11, v22, 22
; SI-NEXT:    v_readlane_b32 s9, v22, 26
; SI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; SI-NEXT:    v_readlane_b32 s99, v20, 35
; SI-NEXT:    v_readlane_b32 s98, v20, 34
; SI-NEXT:    v_readlane_b32 s97, v20, 33
; SI-NEXT:    v_readlane_b32 s96, v20, 32
; SI-NEXT:    v_readlane_b32 s87, v20, 31
; SI-NEXT:    v_readlane_b32 s86, v20, 30
; SI-NEXT:    v_readlane_b32 s85, v20, 29
; SI-NEXT:    v_readlane_b32 s84, v20, 28
; SI-NEXT:    v_readlane_b32 s83, v20, 27
; SI-NEXT:    v_readlane_b32 s82, v20, 26
; SI-NEXT:    v_readlane_b32 s81, v20, 25
; SI-NEXT:    v_readlane_b32 s80, v20, 24
; SI-NEXT:    v_readlane_b32 s71, v20, 23
; SI-NEXT:    v_readlane_b32 s70, v20, 22
; SI-NEXT:    v_readlane_b32 s69, v20, 21
; SI-NEXT:    v_readlane_b32 s68, v20, 20
; SI-NEXT:    v_readlane_b32 s67, v20, 19
; SI-NEXT:    v_readlane_b32 s66, v20, 18
; SI-NEXT:    v_readlane_b32 s65, v20, 17
; SI-NEXT:    v_readlane_b32 s64, v20, 16
; SI-NEXT:    v_readlane_b32 s55, v20, 15
; SI-NEXT:    v_readlane_b32 s54, v20, 14
; SI-NEXT:    v_readlane_b32 s53, v20, 13
; SI-NEXT:    v_readlane_b32 s52, v20, 12
; SI-NEXT:    v_readlane_b32 s51, v20, 11
; SI-NEXT:    v_readlane_b32 s50, v20, 10
; SI-NEXT:    v_readlane_b32 s49, v20, 9
; SI-NEXT:    v_readlane_b32 s48, v20, 8
; SI-NEXT:    v_readlane_b32 s39, v20, 7
; SI-NEXT:    v_readlane_b32 s38, v20, 6
; SI-NEXT:    v_readlane_b32 s37, v20, 5
; SI-NEXT:    v_readlane_b32 s36, v20, 4
; SI-NEXT:    v_readlane_b32 s35, v20, 3
; SI-NEXT:    v_readlane_b32 s34, v20, 2
; SI-NEXT:    v_readlane_b32 s31, v20, 1
; SI-NEXT:    v_readlane_b32 s30, v20, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB99_4:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    v_writelane_b32 v21, s4, 9
; SI-NEXT:    v_writelane_b32 v21, s5, 10
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    v_writelane_b32 v21, s4, 11
; SI-NEXT:    v_writelane_b32 v21, s5, 12
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 13
; SI-NEXT:    v_writelane_b32 v21, s5, 14
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 15
; SI-NEXT:    v_writelane_b32 v21, s5, 16
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 17
; SI-NEXT:    v_writelane_b32 v21, s5, 18
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 19
; SI-NEXT:    v_writelane_b32 v21, s5, 20
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 21
; SI-NEXT:    v_writelane_b32 v21, s5, 22
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 23
; SI-NEXT:    v_writelane_b32 v21, s5, 24
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 25
; SI-NEXT:    v_writelane_b32 v21, s5, 26
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 27
; SI-NEXT:    v_writelane_b32 v21, s5, 28
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 29
; SI-NEXT:    v_writelane_b32 v21, s5, 30
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 31
; SI-NEXT:    v_writelane_b32 v21, s5, 32
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; kill: killed $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 33
; SI-NEXT:    v_writelane_b32 v21, s5, 34
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 35
; SI-NEXT:    v_writelane_b32 v21, s5, 36
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 37
; SI-NEXT:    v_writelane_b32 v21, s5, 38
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 39
; SI-NEXT:    v_writelane_b32 v21, s5, 40
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 41
; SI-NEXT:    v_writelane_b32 v21, s5, 42
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 43
; SI-NEXT:    v_writelane_b32 v21, s5, 44
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 45
; SI-NEXT:    v_writelane_b32 v21, s5, 46
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 47
; SI-NEXT:    v_writelane_b32 v21, s5, 48
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 49
; SI-NEXT:    v_writelane_b32 v21, s5, 50
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 51
; SI-NEXT:    v_writelane_b32 v21, s5, 52
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 53
; SI-NEXT:    v_writelane_b32 v21, s5, 54
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 55
; SI-NEXT:    v_writelane_b32 v21, s5, 56
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 57
; SI-NEXT:    v_writelane_b32 v21, s5, 58
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 59
; SI-NEXT:    v_writelane_b32 v21, s5, 60
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 61
; SI-NEXT:    v_writelane_b32 v21, s5, 62
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v21, s4, 63
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v22, s5, 0
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 1
; SI-NEXT:    v_writelane_b32 v22, s5, 2
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 3
; SI-NEXT:    v_writelane_b32 v22, s5, 4
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 5
; SI-NEXT:    v_writelane_b32 v22, s5, 6
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 7
; SI-NEXT:    v_writelane_b32 v22, s5, 8
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 9
; SI-NEXT:    v_writelane_b32 v22, s5, 10
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 11
; SI-NEXT:    v_writelane_b32 v22, s5, 12
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 13
; SI-NEXT:    v_writelane_b32 v22, s5, 14
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 15
; SI-NEXT:    v_writelane_b32 v22, s5, 16
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 17
; SI-NEXT:    v_writelane_b32 v22, s5, 18
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 19
; SI-NEXT:    v_writelane_b32 v22, s5, 20
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s4, 21
; SI-NEXT:    v_writelane_b32 v22, s5, 22
; SI-NEXT:    v_writelane_b32 v22, s70, 23
; SI-NEXT:    v_writelane_b32 v22, s71, 24
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v22, s70, 25
; SI-NEXT:    v_writelane_b32 v22, s71, 26
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    s_branch .LBB99_2
;
; VI-LABEL: bitcast_v64i16_to_v128i8_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s34, 2
; VI-NEXT:    v_writelane_b32 v32, s35, 3
; VI-NEXT:    v_writelane_b32 v32, s36, 4
; VI-NEXT:    v_writelane_b32 v32, s37, 5
; VI-NEXT:    v_writelane_b32 v32, s38, 6
; VI-NEXT:    v_writelane_b32 v32, s39, 7
; VI-NEXT:    v_writelane_b32 v32, s48, 8
; VI-NEXT:    v_writelane_b32 v32, s49, 9
; VI-NEXT:    v_writelane_b32 v32, s50, 10
; VI-NEXT:    v_writelane_b32 v32, s51, 11
; VI-NEXT:    v_writelane_b32 v32, s52, 12
; VI-NEXT:    v_writelane_b32 v32, s53, 13
; VI-NEXT:    v_writelane_b32 v32, s54, 14
; VI-NEXT:    v_writelane_b32 v32, s55, 15
; VI-NEXT:    v_writelane_b32 v32, s64, 16
; VI-NEXT:    v_writelane_b32 v32, s65, 17
; VI-NEXT:    v_writelane_b32 v32, s66, 18
; VI-NEXT:    v_writelane_b32 v32, s67, 19
; VI-NEXT:    v_writelane_b32 v32, s68, 20
; VI-NEXT:    v_writelane_b32 v32, s69, 21
; VI-NEXT:    v_writelane_b32 v32, s70, 22
; VI-NEXT:    v_writelane_b32 v32, s71, 23
; VI-NEXT:    v_writelane_b32 v32, s80, 24
; VI-NEXT:    v_writelane_b32 v32, s81, 25
; VI-NEXT:    v_writelane_b32 v32, s82, 26
; VI-NEXT:    v_writelane_b32 v32, s83, 27
; VI-NEXT:    v_writelane_b32 v32, s84, 28
; VI-NEXT:    v_writelane_b32 v32, s85, 29
; VI-NEXT:    v_readfirstlane_b32 s44, v19
; VI-NEXT:    v_writelane_b32 v32, s86, 30
; VI-NEXT:    v_readfirstlane_b32 s5, v18
; VI-NEXT:    v_readfirstlane_b32 s4, v17
; VI-NEXT:    v_readfirstlane_b32 s7, v16
; VI-NEXT:    v_readfirstlane_b32 s6, v15
; VI-NEXT:    v_readfirstlane_b32 s9, v14
; VI-NEXT:    v_readfirstlane_b32 s8, v13
; VI-NEXT:    v_readfirstlane_b32 s11, v12
; VI-NEXT:    v_readfirstlane_b32 s10, v11
; VI-NEXT:    v_readfirstlane_b32 s13, v10
; VI-NEXT:    v_readfirstlane_b32 s12, v9
; VI-NEXT:    v_readfirstlane_b32 s15, v8
; VI-NEXT:    v_readfirstlane_b32 s14, v7
; VI-NEXT:    v_readfirstlane_b32 s41, v6
; VI-NEXT:    v_readfirstlane_b32 s40, v5
; VI-NEXT:    v_readfirstlane_b32 s43, v4
; VI-NEXT:    v_readfirstlane_b32 s42, v3
; VI-NEXT:    v_readfirstlane_b32 s45, v2
; VI-NEXT:    s_cmp_lg_u32 s44, 0
; VI-NEXT:    v_readfirstlane_b32 s44, v1
; VI-NEXT:    v_writelane_b32 v32, s87, 31
; VI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; VI-NEXT:    s_cbranch_scc0 .LBB99_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 0
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 1
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s29, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s28, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s27, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s26, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s23, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s23, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s21, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s20, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s19, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s19, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s18, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s17, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s17, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s16, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s14, 16
; VI-NEXT:    s_lshr_b32 s86, s25, 8
; VI-NEXT:    s_lshr_b32 s83, s24, 8
; VI-NEXT:    s_lshr_b32 s84, s23, 8
; VI-NEXT:    s_lshr_b32 s50, s22, 8
; VI-NEXT:    s_lshr_b32 s53, s21, 8
; VI-NEXT:    s_lshr_b32 s65, s20, 8
; VI-NEXT:    s_lshr_b32 s68, s19, 8
; VI-NEXT:    s_lshr_b32 s67, s18, 8
; VI-NEXT:    s_lshr_b32 s69, s17, 8
; VI-NEXT:    s_lshr_b32 s71, s16, 8
; VI-NEXT:    s_lshr_b32 s70, s15, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 57
; VI-NEXT:    s_lshr_b32 s82, s14, 8
; VI-NEXT:    s_lshr_b32 s52, s41, 24
; VI-NEXT:    s_lshr_b32 s80, s41, 16
; VI-NEXT:    s_lshr_b32 s51, s41, 8
; VI-NEXT:    s_lshr_b32 s55, s40, 16
; VI-NEXT:    s_lshr_b32 s85, s40, 8
; VI-NEXT:    s_lshr_b32 s87, s43, 24
; VI-NEXT:    s_lshr_b32 s54, s43, 16
; VI-NEXT:    s_lshr_b32 s66, s43, 8
; VI-NEXT:    s_lshr_b32 s81, s42, 16
; VI-NEXT:    s_lshr_b32 s64, s42, 8
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:    s_cbranch_execnz .LBB99_3
; VI-NEXT:  .LBB99_2: ; %cmp.true
; VI-NEXT:    s_and_b32 s46, s43, 0xffff0000
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_and_b32 s43, s43, 0xffff
; VI-NEXT:    s_or_b32 s43, s46, s43
; VI-NEXT:    s_and_b32 s46, s42, 0xffff0000
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_and_b32 s42, s42, 0xffff
; VI-NEXT:    s_or_b32 s42, s46, s42
; VI-NEXT:    s_and_b32 s46, s41, 0xffff0000
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_and_b32 s41, s41, 0xffff
; VI-NEXT:    s_or_b32 s41, s46, s41
; VI-NEXT:    s_and_b32 s46, s40, 0xffff0000
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_and_b32 s40, s40, 0xffff
; VI-NEXT:    s_or_b32 s40, s46, s40
; VI-NEXT:    s_and_b32 s46, s15, 0xffff0000
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_and_b32 s15, s15, 0xffff
; VI-NEXT:    s_or_b32 s15, s46, s15
; VI-NEXT:    s_and_b32 s46, s14, 0xffff0000
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_and_b32 s14, s14, 0xffff
; VI-NEXT:    s_or_b32 s14, s46, s14
; VI-NEXT:    s_and_b32 s46, s13, 0xffff0000
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_and_b32 s13, s13, 0xffff
; VI-NEXT:    s_or_b32 s13, s46, s13
; VI-NEXT:    s_and_b32 s46, s12, 0xffff0000
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_and_b32 s12, s12, 0xffff
; VI-NEXT:    s_or_b32 s12, s46, s12
; VI-NEXT:    s_and_b32 s46, s11, 0xffff0000
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_and_b32 s11, s11, 0xffff
; VI-NEXT:    s_or_b32 s11, s46, s11
; VI-NEXT:    s_and_b32 s46, s10, 0xffff0000
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_and_b32 s10, s10, 0xffff
; VI-NEXT:    s_or_b32 s10, s46, s10
; VI-NEXT:    s_and_b32 s46, s9, 0xffff0000
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_and_b32 s9, s9, 0xffff
; VI-NEXT:    s_or_b32 s9, s46, s9
; VI-NEXT:    s_and_b32 s46, s8, 0xffff0000
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_and_b32 s8, s8, 0xffff
; VI-NEXT:    s_or_b32 s8, s46, s8
; VI-NEXT:    s_and_b32 s46, s7, 0xffff0000
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_and_b32 s7, s7, 0xffff
; VI-NEXT:    s_or_b32 s7, s46, s7
; VI-NEXT:    s_and_b32 s46, s6, 0xffff0000
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_and_b32 s6, s6, 0xffff
; VI-NEXT:    s_or_b32 s6, s46, s6
; VI-NEXT:    s_and_b32 s46, s5, 0xffff0000
; VI-NEXT:    s_add_i32 s5, s5, 3
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s5, s46, s5
; VI-NEXT:    s_and_b32 s46, s4, 0xffff0000
; VI-NEXT:    s_add_i32 s4, s4, 3
; VI-NEXT:    s_and_b32 s4, s4, 0xffff
; VI-NEXT:    s_or_b32 s4, s46, s4
; VI-NEXT:    s_and_b32 s46, s17, 0xffff0000
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_and_b32 s17, s17, 0xffff
; VI-NEXT:    s_or_b32 s17, s46, s17
; VI-NEXT:    s_and_b32 s46, s16, 0xffff0000
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    s_and_b32 s16, s16, 0xffff
; VI-NEXT:    s_or_b32 s16, s46, s16
; VI-NEXT:    s_and_b32 s46, s19, 0xffff0000
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_and_b32 s19, s19, 0xffff
; VI-NEXT:    s_or_b32 s19, s46, s19
; VI-NEXT:    s_and_b32 s46, s18, 0xffff0000
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_and_b32 s18, s18, 0xffff
; VI-NEXT:    s_or_b32 s18, s46, s18
; VI-NEXT:    s_and_b32 s46, s21, 0xffff0000
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_and_b32 s21, s21, 0xffff
; VI-NEXT:    s_or_b32 s21, s46, s21
; VI-NEXT:    s_and_b32 s46, s20, 0xffff0000
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_and_b32 s20, s20, 0xffff
; VI-NEXT:    s_or_b32 s20, s46, s20
; VI-NEXT:    s_and_b32 s46, s23, 0xffff0000
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_and_b32 s23, s23, 0xffff
; VI-NEXT:    s_or_b32 s23, s46, s23
; VI-NEXT:    s_and_b32 s46, s22, 0xffff0000
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_and_b32 s22, s22, 0xffff
; VI-NEXT:    s_or_b32 s22, s46, s22
; VI-NEXT:    s_and_b32 s46, s25, 0xffff0000
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_and_b32 s25, s25, 0xffff
; VI-NEXT:    s_or_b32 s25, s46, s25
; VI-NEXT:    s_and_b32 s46, s24, 0xffff0000
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_and_b32 s24, s24, 0xffff
; VI-NEXT:    s_or_b32 s24, s46, s24
; VI-NEXT:    s_and_b32 s46, s27, 0xffff0000
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_and_b32 s27, s27, 0xffff
; VI-NEXT:    s_or_b32 s27, s46, s27
; VI-NEXT:    s_and_b32 s46, s26, 0xffff0000
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_and_b32 s26, s26, 0xffff
; VI-NEXT:    s_or_b32 s26, s46, s26
; VI-NEXT:    s_and_b32 s46, s29, 0xffff0000
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_and_b32 s29, s29, 0xffff
; VI-NEXT:    s_or_b32 s29, s46, s29
; VI-NEXT:    s_and_b32 s46, s28, 0xffff0000
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_and_b32 s28, s28, 0xffff
; VI-NEXT:    s_or_b32 s28, s46, s28
; VI-NEXT:    s_and_b32 s46, s45, 0xffff0000
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:    s_and_b32 s45, s45, 0xffff
; VI-NEXT:    s_or_b32 s45, s46, s45
; VI-NEXT:    s_and_b32 s46, s44, 0xffff0000
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_add_i32 s45, s45, 0x30000
; VI-NEXT:    s_and_b32 s44, s44, 0xffff
; VI-NEXT:    s_or_b32 s44, s46, s44
; VI-NEXT:    s_lshr_b32 s46, s45, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 0
; VI-NEXT:    s_lshr_b32 s46, s45, 16
; VI-NEXT:    s_add_i32 s44, s44, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 1
; VI-NEXT:    s_lshr_b32 s46, s45, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 2
; VI-NEXT:    s_lshr_b32 s46, s44, 16
; VI-NEXT:    s_add_i32 s29, s29, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 3
; VI-NEXT:    s_lshr_b32 s46, s44, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 4
; VI-NEXT:    s_lshr_b32 s46, s29, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 5
; VI-NEXT:    s_lshr_b32 s46, s29, 16
; VI-NEXT:    s_add_i32 s28, s28, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 6
; VI-NEXT:    s_lshr_b32 s46, s29, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 7
; VI-NEXT:    s_lshr_b32 s46, s28, 16
; VI-NEXT:    s_add_i32 s27, s27, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 8
; VI-NEXT:    s_lshr_b32 s46, s28, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 9
; VI-NEXT:    s_lshr_b32 s46, s27, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 10
; VI-NEXT:    s_lshr_b32 s46, s27, 16
; VI-NEXT:    s_add_i32 s26, s26, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 11
; VI-NEXT:    s_lshr_b32 s46, s27, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 12
; VI-NEXT:    s_lshr_b32 s46, s26, 16
; VI-NEXT:    s_add_i32 s25, s25, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 13
; VI-NEXT:    s_lshr_b32 s46, s26, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 14
; VI-NEXT:    s_lshr_b32 s46, s25, 24
; VI-NEXT:    s_add_i32 s24, s24, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 15
; VI-NEXT:    s_lshr_b32 s46, s25, 16
; VI-NEXT:    s_add_i32 s23, s23, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 16
; VI-NEXT:    s_lshr_b32 s46, s24, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 17
; VI-NEXT:    s_lshr_b32 s46, s23, 24
; VI-NEXT:    s_add_i32 s22, s22, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 18
; VI-NEXT:    s_lshr_b32 s46, s23, 16
; VI-NEXT:    s_add_i32 s21, s21, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 19
; VI-NEXT:    s_lshr_b32 s46, s22, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 20
; VI-NEXT:    s_lshr_b32 s46, s21, 24
; VI-NEXT:    s_add_i32 s20, s20, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 21
; VI-NEXT:    s_lshr_b32 s46, s21, 16
; VI-NEXT:    s_add_i32 s19, s19, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 22
; VI-NEXT:    s_lshr_b32 s46, s20, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 23
; VI-NEXT:    s_lshr_b32 s46, s19, 24
; VI-NEXT:    s_add_i32 s18, s18, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 24
; VI-NEXT:    s_lshr_b32 s46, s19, 16
; VI-NEXT:    s_add_i32 s17, s17, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 25
; VI-NEXT:    s_lshr_b32 s46, s18, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 26
; VI-NEXT:    s_lshr_b32 s46, s17, 24
; VI-NEXT:    s_add_i32 s16, s16, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 27
; VI-NEXT:    s_lshr_b32 s46, s17, 16
; VI-NEXT:    s_add_i32 s5, s5, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 28
; VI-NEXT:    s_lshr_b32 s46, s16, 16
; VI-NEXT:    v_writelane_b32 v33, s46, 29
; VI-NEXT:    s_lshr_b32 s46, s5, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 30
; VI-NEXT:    s_lshr_b32 s46, s5, 16
; VI-NEXT:    s_add_i32 s4, s4, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 31
; VI-NEXT:    s_lshr_b32 s46, s5, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 32
; VI-NEXT:    s_lshr_b32 s46, s4, 16
; VI-NEXT:    s_add_i32 s7, s7, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 33
; VI-NEXT:    s_lshr_b32 s46, s4, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 34
; VI-NEXT:    s_lshr_b32 s46, s7, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 35
; VI-NEXT:    s_lshr_b32 s46, s7, 16
; VI-NEXT:    s_add_i32 s6, s6, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 36
; VI-NEXT:    s_lshr_b32 s46, s7, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 37
; VI-NEXT:    s_lshr_b32 s46, s6, 16
; VI-NEXT:    s_add_i32 s9, s9, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 38
; VI-NEXT:    s_lshr_b32 s46, s6, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 39
; VI-NEXT:    s_lshr_b32 s46, s9, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 40
; VI-NEXT:    s_lshr_b32 s46, s9, 16
; VI-NEXT:    s_add_i32 s8, s8, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 41
; VI-NEXT:    s_lshr_b32 s46, s9, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 42
; VI-NEXT:    s_lshr_b32 s46, s8, 16
; VI-NEXT:    s_add_i32 s11, s11, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 43
; VI-NEXT:    s_lshr_b32 s46, s8, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 44
; VI-NEXT:    s_lshr_b32 s46, s11, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 45
; VI-NEXT:    s_lshr_b32 s46, s11, 16
; VI-NEXT:    s_add_i32 s10, s10, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 46
; VI-NEXT:    s_lshr_b32 s46, s11, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 47
; VI-NEXT:    s_lshr_b32 s46, s10, 16
; VI-NEXT:    s_add_i32 s13, s13, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 48
; VI-NEXT:    s_lshr_b32 s46, s10, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 49
; VI-NEXT:    s_lshr_b32 s46, s13, 24
; VI-NEXT:    v_writelane_b32 v33, s46, 50
; VI-NEXT:    s_lshr_b32 s46, s13, 16
; VI-NEXT:    s_add_i32 s12, s12, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 51
; VI-NEXT:    s_lshr_b32 s46, s13, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 52
; VI-NEXT:    s_lshr_b32 s46, s12, 16
; VI-NEXT:    s_add_i32 s15, s15, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 53
; VI-NEXT:    s_lshr_b32 s46, s12, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 54
; VI-NEXT:    s_lshr_b32 s46, s15, 24
; VI-NEXT:    s_add_i32 s14, s14, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 55
; VI-NEXT:    s_lshr_b32 s46, s15, 16
; VI-NEXT:    s_add_i32 s43, s43, 0x30000
; VI-NEXT:    s_add_i32 s42, s42, 0x30000
; VI-NEXT:    s_add_i32 s41, s41, 0x30000
; VI-NEXT:    s_add_i32 s40, s40, 0x30000
; VI-NEXT:    v_writelane_b32 v33, s46, 56
; VI-NEXT:    s_lshr_b32 s46, s14, 16
; VI-NEXT:    s_lshr_b32 s86, s25, 8
; VI-NEXT:    s_lshr_b32 s83, s24, 8
; VI-NEXT:    s_lshr_b32 s84, s23, 8
; VI-NEXT:    s_lshr_b32 s50, s22, 8
; VI-NEXT:    s_lshr_b32 s53, s21, 8
; VI-NEXT:    s_lshr_b32 s65, s20, 8
; VI-NEXT:    s_lshr_b32 s68, s19, 8
; VI-NEXT:    s_lshr_b32 s67, s18, 8
; VI-NEXT:    s_lshr_b32 s69, s17, 8
; VI-NEXT:    s_lshr_b32 s71, s16, 8
; VI-NEXT:    s_lshr_b32 s70, s15, 8
; VI-NEXT:    v_writelane_b32 v33, s46, 57
; VI-NEXT:    s_lshr_b32 s82, s14, 8
; VI-NEXT:    s_lshr_b32 s52, s41, 24
; VI-NEXT:    s_lshr_b32 s80, s41, 16
; VI-NEXT:    s_lshr_b32 s51, s41, 8
; VI-NEXT:    s_lshr_b32 s55, s40, 16
; VI-NEXT:    s_lshr_b32 s85, s40, 8
; VI-NEXT:    s_lshr_b32 s87, s43, 24
; VI-NEXT:    s_lshr_b32 s54, s43, 16
; VI-NEXT:    s_lshr_b32 s66, s43, 8
; VI-NEXT:    s_lshr_b32 s81, s42, 16
; VI-NEXT:    s_lshr_b32 s64, s42, 8
; VI-NEXT:    s_lshr_b64 s[78:79], s[44:45], 24
; VI-NEXT:    s_lshr_b64 s[88:89], s[28:29], 24
; VI-NEXT:    s_lshr_b64 s[90:91], s[26:27], 24
; VI-NEXT:    s_lshr_b64 s[30:31], s[24:25], 24
; VI-NEXT:    s_lshr_b64 s[34:35], s[22:23], 24
; VI-NEXT:    s_lshr_b64 s[36:37], s[20:21], 24
; VI-NEXT:    s_lshr_b64 s[38:39], s[18:19], 24
; VI-NEXT:    s_lshr_b64 s[48:49], s[16:17], 24
; VI-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; VI-NEXT:    s_lshr_b64 s[56:57], s[6:7], 24
; VI-NEXT:    s_lshr_b64 s[58:59], s[8:9], 24
; VI-NEXT:    s_lshr_b64 s[60:61], s[10:11], 24
; VI-NEXT:    s_lshr_b64 s[62:63], s[12:13], 24
; VI-NEXT:    s_lshr_b64 s[72:73], s[14:15], 24
; VI-NEXT:    s_lshr_b64 s[74:75], s[40:41], 24
; VI-NEXT:    s_lshr_b64 s[76:77], s[42:43], 24
; VI-NEXT:  .LBB99_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v2, s71
; VI-NEXT:    v_mov_b32_e32 v4, s69
; VI-NEXT:    v_mov_b32_e32 v1, 0xc0c0004
; VI-NEXT:    v_perm_b32 v3, s16, v2, v1
; VI-NEXT:    v_perm_b32 v2, s17, v4, v1
; VI-NEXT:    v_mov_b32_e32 v4, s67
; VI-NEXT:    v_mov_b32_e32 v6, s68
; VI-NEXT:    v_perm_b32 v5, s18, v4, v1
; VI-NEXT:    v_perm_b32 v4, s19, v6, v1
; VI-NEXT:    v_mov_b32_e32 v6, s65
; VI-NEXT:    v_mov_b32_e32 v8, s53
; VI-NEXT:    v_perm_b32 v7, s20, v6, v1
; VI-NEXT:    v_perm_b32 v6, s21, v8, v1
; VI-NEXT:    v_mov_b32_e32 v8, s50
; VI-NEXT:    v_mov_b32_e32 v11, s84
; VI-NEXT:    v_perm_b32 v9, s22, v8, v1
; VI-NEXT:    v_perm_b32 v8, s23, v11, v1
; VI-NEXT:    v_mov_b32_e32 v11, s83
; VI-NEXT:    v_mov_b32_e32 v14, s86
; VI-NEXT:    v_readlane_b32 s16, v33, 14
; VI-NEXT:    v_perm_b32 v12, s24, v11, v1
; VI-NEXT:    v_perm_b32 v11, s25, v14, v1
; VI-NEXT:    v_mov_b32_e32 v14, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 12
; VI-NEXT:    v_mov_b32_e32 v17, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 9
; VI-NEXT:    v_perm_b32 v15, s26, v14, v1
; VI-NEXT:    v_perm_b32 v14, s27, v17, v1
; VI-NEXT:    v_mov_b32_e32 v17, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 7
; VI-NEXT:    v_mov_b32_e32 v23, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 4
; VI-NEXT:    v_mov_b32_e32 v24, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 2
; VI-NEXT:    v_mov_b32_e32 v10, s48
; VI-NEXT:    v_mov_b32_e32 v25, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 29
; VI-NEXT:    v_perm_b32 v10, s16, v10, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 27
; VI-NEXT:    v_mov_b32_e32 v26, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 28
; VI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; VI-NEXT:    v_perm_b32 v26, s16, v26, v1
; VI-NEXT:    v_or_b32_e32 v3, v3, v10
; VI-NEXT:    v_mov_b32_e32 v13, s38
; VI-NEXT:    v_readlane_b32 s16, v33, 26
; VI-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v26
; VI-NEXT:    v_perm_b32 v13, s16, v13, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 24
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 4, v0
; VI-NEXT:    v_mov_b32_e32 v27, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 25
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v13
; VI-NEXT:    v_perm_b32 v27, s16, v27, v1
; VI-NEXT:    v_or_b32_e32 v2, v5, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 8, v0
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_readlane_b32 s16, v33, 23
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v27
; VI-NEXT:    v_perm_b32 v16, s16, v16, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 21
; VI-NEXT:    v_or_b32_e32 v2, v4, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 12, v0
; VI-NEXT:    v_mov_b32_e32 v28, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 22
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v16
; VI-NEXT:    v_perm_b32 v28, s16, v28, v1
; VI-NEXT:    v_or_b32_e32 v2, v7, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 16, v0
; VI-NEXT:    v_mov_b32_e32 v19, s34
; VI-NEXT:    v_readlane_b32 s16, v33, 20
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v28
; VI-NEXT:    v_perm_b32 v19, s16, v19, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 18
; VI-NEXT:    v_or_b32_e32 v2, v6, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 20, v0
; VI-NEXT:    v_mov_b32_e32 v29, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 19
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v19
; VI-NEXT:    v_perm_b32 v29, s16, v29, v1
; VI-NEXT:    v_or_b32_e32 v2, v9, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 24, v0
; VI-NEXT:    v_mov_b32_e32 v20, s30
; VI-NEXT:    v_readlane_b32 s16, v33, 17
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v29
; VI-NEXT:    v_perm_b32 v20, s16, v20, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 15
; VI-NEXT:    v_or_b32_e32 v2, v8, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 28, v0
; VI-NEXT:    v_mov_b32_e32 v30, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 16
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v20
; VI-NEXT:    v_perm_b32 v30, s16, v30, v1
; VI-NEXT:    v_or_b32_e32 v2, v12, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 32, v0
; VI-NEXT:    v_mov_b32_e32 v21, s90
; VI-NEXT:    v_readlane_b32 s16, v33, 13
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v30
; VI-NEXT:    v_perm_b32 v21, s16, v21, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 10
; VI-NEXT:    v_or_b32_e32 v2, v11, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 36, v0
; VI-NEXT:    v_mov_b32_e32 v31, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 11
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v21
; VI-NEXT:    v_perm_b32 v31, s16, v31, v1
; VI-NEXT:    v_or_b32_e32 v2, v15, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 40, v0
; VI-NEXT:    v_mov_b32_e32 v22, s88
; VI-NEXT:    v_readlane_b32 s16, v33, 8
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v31
; VI-NEXT:    v_perm_b32 v22, s16, v22, v1
; VI-NEXT:    v_readlane_b32 s16, v33, 5
; VI-NEXT:    v_or_b32_e32 v2, v14, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 44, v0
; VI-NEXT:    v_perm_b32 v18, s28, v17, v1
; VI-NEXT:    v_mov_b32_e32 v10, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 6
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v22
; VI-NEXT:    v_perm_b32 v3, s16, v10, v1
; VI-NEXT:    v_or_b32_e32 v2, v18, v2
; VI-NEXT:    v_add_u32_e32 v4, vcc, 48, v0
; VI-NEXT:    v_perm_b32 v17, s29, v23, v1
; VI-NEXT:    buffer_store_dword v2, v4, s[0:3], 0 offen
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v3
; VI-NEXT:    v_mov_b32_e32 v23, s78
; VI-NEXT:    v_or_b32_e32 v2, v17, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 52, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 3
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_perm_b32 v2, s16, v23, v1
; VI-NEXT:    v_perm_b32 v24, s44, v24, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v2, v24, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 56, v0
; VI-NEXT:    v_readlane_b32 s16, v33, 0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s16
; VI-NEXT:    v_readlane_b32 s16, v33, 1
; VI-NEXT:    v_perm_b32 v2, s16, v2, v1
; VI-NEXT:    v_perm_b32 v25, s45, v25, v1
; VI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; VI-NEXT:    v_or_b32_e32 v2, v25, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 60, v0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v3, s76
; VI-NEXT:    v_mov_b32_e32 v2, s64
; VI-NEXT:    v_perm_b32 v3, s81, v3, v1
; VI-NEXT:    v_perm_b32 v2, s42, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 64, v0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v3, s87
; VI-NEXT:    v_mov_b32_e32 v2, s66
; VI-NEXT:    v_perm_b32 v3, s54, v3, v1
; VI-NEXT:    v_perm_b32 v2, s43, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x44, v0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v3, s74
; VI-NEXT:    v_mov_b32_e32 v2, s85
; VI-NEXT:    v_perm_b32 v3, s55, v3, v1
; VI-NEXT:    v_perm_b32 v2, s40, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x48, v0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v3, s52
; VI-NEXT:    v_mov_b32_e32 v2, s51
; VI-NEXT:    v_perm_b32 v3, s80, v3, v1
; VI-NEXT:    v_perm_b32 v2, s41, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x4c, v0
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s82
; VI-NEXT:    v_perm_b32 v2, s14, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s72
; VI-NEXT:    v_readlane_b32 s14, v33, 57
; VI-NEXT:    v_perm_b32 v3, s14, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x50, v0
; VI-NEXT:    v_readlane_b32 s14, v33, 55
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v3, s14
; VI-NEXT:    v_readlane_b32 s14, v33, 56
; VI-NEXT:    v_mov_b32_e32 v2, s70
; VI-NEXT:    v_perm_b32 v3, s14, v3, v1
; VI-NEXT:    v_perm_b32 v2, s15, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x54, v0
; VI-NEXT:    v_readlane_b32 s14, v33, 54
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s14
; VI-NEXT:    v_perm_b32 v2, s12, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s62
; VI-NEXT:    v_readlane_b32 s12, v33, 53
; VI-NEXT:    v_perm_b32 v3, s12, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x58, v0
; VI-NEXT:    v_readlane_b32 s12, v33, 52
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s12
; VI-NEXT:    v_readlane_b32 s12, v33, 50
; VI-NEXT:    v_mov_b32_e32 v3, s12
; VI-NEXT:    v_readlane_b32 s12, v33, 51
; VI-NEXT:    v_perm_b32 v3, s12, v3, v1
; VI-NEXT:    v_perm_b32 v2, s13, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x5c, v0
; VI-NEXT:    v_readlane_b32 s12, v33, 49
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s12
; VI-NEXT:    v_perm_b32 v2, s10, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s60
; VI-NEXT:    v_readlane_b32 s10, v33, 48
; VI-NEXT:    v_perm_b32 v3, s10, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x60, v0
; VI-NEXT:    v_readlane_b32 s10, v33, 47
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s10
; VI-NEXT:    v_readlane_b32 s10, v33, 45
; VI-NEXT:    v_mov_b32_e32 v3, s10
; VI-NEXT:    v_readlane_b32 s10, v33, 46
; VI-NEXT:    v_perm_b32 v3, s10, v3, v1
; VI-NEXT:    v_perm_b32 v2, s11, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x64, v0
; VI-NEXT:    v_readlane_b32 s10, v33, 44
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s10
; VI-NEXT:    v_perm_b32 v2, s8, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s58
; VI-NEXT:    v_readlane_b32 s8, v33, 43
; VI-NEXT:    v_perm_b32 v3, s8, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x68, v0
; VI-NEXT:    v_readlane_b32 s8, v33, 42
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s8
; VI-NEXT:    v_readlane_b32 s8, v33, 40
; VI-NEXT:    v_mov_b32_e32 v3, s8
; VI-NEXT:    v_readlane_b32 s8, v33, 41
; VI-NEXT:    v_perm_b32 v3, s8, v3, v1
; VI-NEXT:    v_perm_b32 v2, s9, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x6c, v0
; VI-NEXT:    v_readlane_b32 s8, v33, 39
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s8
; VI-NEXT:    v_perm_b32 v2, s6, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s56
; VI-NEXT:    v_readlane_b32 s6, v33, 38
; VI-NEXT:    v_perm_b32 v3, s6, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x70, v0
; VI-NEXT:    v_readlane_b32 s6, v33, 37
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 35
; VI-NEXT:    v_mov_b32_e32 v3, s6
; VI-NEXT:    v_readlane_b32 s6, v33, 36
; VI-NEXT:    v_perm_b32 v3, s6, v3, v1
; VI-NEXT:    v_perm_b32 v2, s7, v2, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x74, v0
; VI-NEXT:    v_readlane_b32 s6, v33, 34
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s6
; VI-NEXT:    v_perm_b32 v2, s4, v2, v1
; VI-NEXT:    v_mov_b32_e32 v3, s46
; VI-NEXT:    v_readlane_b32 s4, v33, 33
; VI-NEXT:    v_perm_b32 v3, s4, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_or_b32_e32 v2, v2, v3
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x78, v0
; VI-NEXT:    v_readlane_b32 s4, v33, 32
; VI-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
; VI-NEXT:    v_mov_b32_e32 v2, s4
; VI-NEXT:    v_readlane_b32 s4, v33, 30
; VI-NEXT:    v_mov_b32_e32 v3, s4
; VI-NEXT:    v_readlane_b32 s4, v33, 31
; VI-NEXT:    v_perm_b32 v2, s5, v2, v1
; VI-NEXT:    v_perm_b32 v1, s4, v3, v1
; VI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_or_b32_e32 v1, v2, v1
; VI-NEXT:    v_add_u32_e32 v0, vcc, 0x7c, v0
; VI-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
; VI-NEXT:    v_readlane_b32 s87, v32, 31
; VI-NEXT:    v_readlane_b32 s86, v32, 30
; VI-NEXT:    v_readlane_b32 s85, v32, 29
; VI-NEXT:    v_readlane_b32 s84, v32, 28
; VI-NEXT:    v_readlane_b32 s83, v32, 27
; VI-NEXT:    v_readlane_b32 s82, v32, 26
; VI-NEXT:    v_readlane_b32 s81, v32, 25
; VI-NEXT:    v_readlane_b32 s80, v32, 24
; VI-NEXT:    v_readlane_b32 s71, v32, 23
; VI-NEXT:    v_readlane_b32 s70, v32, 22
; VI-NEXT:    v_readlane_b32 s69, v32, 21
; VI-NEXT:    v_readlane_b32 s68, v32, 20
; VI-NEXT:    v_readlane_b32 s67, v32, 19
; VI-NEXT:    v_readlane_b32 s66, v32, 18
; VI-NEXT:    v_readlane_b32 s65, v32, 17
; VI-NEXT:    v_readlane_b32 s64, v32, 16
; VI-NEXT:    v_readlane_b32 s55, v32, 15
; VI-NEXT:    v_readlane_b32 s54, v32, 14
; VI-NEXT:    v_readlane_b32 s53, v32, 13
; VI-NEXT:    v_readlane_b32 s52, v32, 12
; VI-NEXT:    v_readlane_b32 s51, v32, 11
; VI-NEXT:    v_readlane_b32 s50, v32, 10
; VI-NEXT:    v_readlane_b32 s49, v32, 9
; VI-NEXT:    v_readlane_b32 s48, v32, 8
; VI-NEXT:    v_readlane_b32 s39, v32, 7
; VI-NEXT:    v_readlane_b32 s38, v32, 6
; VI-NEXT:    v_readlane_b32 s37, v32, 5
; VI-NEXT:    v_readlane_b32 s36, v32, 4
; VI-NEXT:    v_readlane_b32 s35, v32, 3
; VI-NEXT:    v_readlane_b32 s34, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB99_4:
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr71
; VI-NEXT:    ; implicit-def: $sgpr69
; VI-NEXT:    ; implicit-def: $sgpr67
; VI-NEXT:    ; implicit-def: $sgpr68
; VI-NEXT:    ; implicit-def: $sgpr65
; VI-NEXT:    ; implicit-def: $sgpr53
; VI-NEXT:    ; implicit-def: $sgpr50
; VI-NEXT:    ; implicit-def: $sgpr84
; VI-NEXT:    ; implicit-def: $sgpr83
; VI-NEXT:    ; implicit-def: $sgpr86
; VI-NEXT:    ; implicit-def: $sgpr64
; VI-NEXT:    ; implicit-def: $sgpr81
; VI-NEXT:    ; implicit-def: $sgpr66
; VI-NEXT:    ; implicit-def: $sgpr54
; VI-NEXT:    ; implicit-def: $sgpr87
; VI-NEXT:    ; implicit-def: $sgpr85
; VI-NEXT:    ; implicit-def: $sgpr55
; VI-NEXT:    ; implicit-def: $sgpr51
; VI-NEXT:    ; implicit-def: $sgpr80
; VI-NEXT:    ; implicit-def: $sgpr52
; VI-NEXT:    ; implicit-def: $sgpr82
; VI-NEXT:    ; implicit-def: $sgpr70
; VI-NEXT:    ; implicit-def: $sgpr48
; VI-NEXT:    ; implicit-def: $sgpr38
; VI-NEXT:    ; implicit-def: $sgpr36
; VI-NEXT:    ; implicit-def: $sgpr34
; VI-NEXT:    ; implicit-def: $sgpr30
; VI-NEXT:    ; implicit-def: $sgpr90
; VI-NEXT:    ; implicit-def: $sgpr88
; VI-NEXT:    ; implicit-def: $sgpr78
; VI-NEXT:    ; implicit-def: $sgpr76
; VI-NEXT:    ; implicit-def: $sgpr74
; VI-NEXT:    ; implicit-def: $sgpr72
; VI-NEXT:    ; implicit-def: $sgpr62
; VI-NEXT:    ; implicit-def: $sgpr60
; VI-NEXT:    ; implicit-def: $sgpr58
; VI-NEXT:    ; implicit-def: $sgpr56
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    ; kill: killed $sgpr46
; VI-NEXT:    ; implicit-def: $sgpr46
; VI-NEXT:    s_branch .LBB99_2
;
; GFX9-LABEL: bitcast_v64i16_to_v128i8_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v63, s30, 0
; GFX9-NEXT:    v_writelane_b32 v63, s31, 1
; GFX9-NEXT:    v_writelane_b32 v63, s34, 2
; GFX9-NEXT:    v_writelane_b32 v63, s35, 3
; GFX9-NEXT:    v_writelane_b32 v63, s36, 4
; GFX9-NEXT:    v_writelane_b32 v63, s37, 5
; GFX9-NEXT:    v_writelane_b32 v63, s38, 6
; GFX9-NEXT:    v_writelane_b32 v63, s39, 7
; GFX9-NEXT:    v_writelane_b32 v63, s48, 8
; GFX9-NEXT:    v_writelane_b32 v63, s49, 9
; GFX9-NEXT:    v_writelane_b32 v63, s50, 10
; GFX9-NEXT:    v_writelane_b32 v63, s51, 11
; GFX9-NEXT:    v_writelane_b32 v63, s52, 12
; GFX9-NEXT:    v_writelane_b32 v63, s53, 13
; GFX9-NEXT:    v_writelane_b32 v63, s54, 14
; GFX9-NEXT:    v_writelane_b32 v63, s55, 15
; GFX9-NEXT:    v_writelane_b32 v63, s64, 16
; GFX9-NEXT:    v_writelane_b32 v63, s65, 17
; GFX9-NEXT:    v_writelane_b32 v63, s66, 18
; GFX9-NEXT:    v_writelane_b32 v63, s67, 19
; GFX9-NEXT:    v_writelane_b32 v63, s68, 20
; GFX9-NEXT:    v_writelane_b32 v63, s69, 21
; GFX9-NEXT:    v_writelane_b32 v63, s70, 22
; GFX9-NEXT:    v_writelane_b32 v63, s71, 23
; GFX9-NEXT:    v_writelane_b32 v63, s80, 24
; GFX9-NEXT:    v_writelane_b32 v63, s81, 25
; GFX9-NEXT:    v_writelane_b32 v63, s82, 26
; GFX9-NEXT:    v_writelane_b32 v63, s83, 27
; GFX9-NEXT:    v_writelane_b32 v63, s84, 28
; GFX9-NEXT:    v_writelane_b32 v63, s85, 29
; GFX9-NEXT:    v_writelane_b32 v63, s86, 30
; GFX9-NEXT:    v_writelane_b32 v63, s87, 31
; GFX9-NEXT:    v_writelane_b32 v63, s96, 32
; GFX9-NEXT:    v_writelane_b32 v63, s97, 33
; GFX9-NEXT:    v_writelane_b32 v63, s98, 34
; GFX9-NEXT:    v_readfirstlane_b32 s4, v19
; GFX9-NEXT:    v_writelane_b32 v63, s99, 35
; GFX9-NEXT:    v_readfirstlane_b32 s7, v18
; GFX9-NEXT:    v_readfirstlane_b32 s6, v17
; GFX9-NEXT:    v_readfirstlane_b32 s9, v16
; GFX9-NEXT:    v_readfirstlane_b32 s8, v15
; GFX9-NEXT:    v_readfirstlane_b32 s11, v14
; GFX9-NEXT:    v_readfirstlane_b32 s10, v13
; GFX9-NEXT:    v_readfirstlane_b32 s13, v12
; GFX9-NEXT:    v_readfirstlane_b32 s12, v11
; GFX9-NEXT:    v_readfirstlane_b32 s15, v10
; GFX9-NEXT:    v_readfirstlane_b32 s14, v9
; GFX9-NEXT:    v_readfirstlane_b32 s41, v8
; GFX9-NEXT:    v_readfirstlane_b32 s40, v7
; GFX9-NEXT:    v_readfirstlane_b32 s43, v6
; GFX9-NEXT:    v_readfirstlane_b32 s42, v5
; GFX9-NEXT:    v_readfirstlane_b32 s45, v4
; GFX9-NEXT:    v_readfirstlane_b32 s44, v3
; GFX9-NEXT:    v_readfirstlane_b32 s5, v2
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s4, v1
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; GFX9-NEXT:    s_cbranch_scc0 .LBB99_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_lshr_b32 s46, s5, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 49
; GFX9-NEXT:    s_lshr_b32 s46, s5, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 48
; GFX9-NEXT:    s_lshr_b32 s46, s5, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 47
; GFX9-NEXT:    s_lshr_b32 s46, s4, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 46
; GFX9-NEXT:    s_lshr_b32 s46, s4, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 45
; GFX9-NEXT:    s_lshr_b32 s46, s29, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 44
; GFX9-NEXT:    s_lshr_b32 s46, s29, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 43
; GFX9-NEXT:    s_lshr_b32 s46, s29, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 42
; GFX9-NEXT:    s_lshr_b32 s46, s28, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 41
; GFX9-NEXT:    s_lshr_b32 s46, s28, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 40
; GFX9-NEXT:    s_lshr_b32 s46, s27, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 39
; GFX9-NEXT:    s_lshr_b32 s46, s27, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 38
; GFX9-NEXT:    s_lshr_b32 s46, s27, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 37
; GFX9-NEXT:    s_lshr_b32 s46, s26, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 36
; GFX9-NEXT:    s_lshr_b32 s46, s26, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 35
; GFX9-NEXT:    s_lshr_b32 s46, s25, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 34
; GFX9-NEXT:    s_lshr_b32 s46, s25, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 33
; GFX9-NEXT:    s_lshr_b32 s46, s25, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 32
; GFX9-NEXT:    s_lshr_b32 s46, s24, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 31
; GFX9-NEXT:    s_lshr_b32 s46, s24, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 30
; GFX9-NEXT:    s_lshr_b32 s46, s23, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 29
; GFX9-NEXT:    s_lshr_b32 s46, s23, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 28
; GFX9-NEXT:    s_lshr_b32 s46, s23, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 27
; GFX9-NEXT:    s_lshr_b32 s46, s22, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 26
; GFX9-NEXT:    s_lshr_b32 s46, s22, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 25
; GFX9-NEXT:    s_lshr_b32 s46, s21, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 24
; GFX9-NEXT:    s_lshr_b32 s46, s21, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 23
; GFX9-NEXT:    s_lshr_b32 s46, s21, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 22
; GFX9-NEXT:    s_lshr_b32 s46, s20, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 21
; GFX9-NEXT:    s_lshr_b32 s46, s20, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 20
; GFX9-NEXT:    s_lshr_b32 s46, s19, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 19
; GFX9-NEXT:    s_lshr_b32 s46, s19, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 18
; GFX9-NEXT:    s_lshr_b32 s46, s19, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 17
; GFX9-NEXT:    s_lshr_b32 s46, s18, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 16
; GFX9-NEXT:    s_lshr_b32 s46, s18, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 15
; GFX9-NEXT:    s_lshr_b32 s46, s17, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 14
; GFX9-NEXT:    s_lshr_b32 s46, s17, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 13
; GFX9-NEXT:    s_lshr_b32 s46, s17, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 12
; GFX9-NEXT:    s_lshr_b32 s46, s16, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 11
; GFX9-NEXT:    s_lshr_b32 s46, s16, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 10
; GFX9-NEXT:    s_lshr_b32 s46, s7, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 9
; GFX9-NEXT:    s_lshr_b32 s46, s7, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 8
; GFX9-NEXT:    s_lshr_b32 s46, s7, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 7
; GFX9-NEXT:    s_lshr_b32 s46, s6, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 6
; GFX9-NEXT:    s_lshr_b32 s46, s6, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 5
; GFX9-NEXT:    s_lshr_b32 s46, s9, 24
; GFX9-NEXT:    v_writelane_b32 v62, s46, 4
; GFX9-NEXT:    s_lshr_b32 s46, s9, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 3
; GFX9-NEXT:    s_lshr_b32 s46, s9, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 2
; GFX9-NEXT:    s_lshr_b32 s46, s8, 16
; GFX9-NEXT:    v_writelane_b32 v62, s46, 1
; GFX9-NEXT:    s_lshr_b32 s46, s8, 8
; GFX9-NEXT:    v_writelane_b32 v62, s46, 0
; GFX9-NEXT:    s_lshr_b32 s82, s11, 24
; GFX9-NEXT:    s_lshr_b32 s83, s11, 16
; GFX9-NEXT:    s_lshr_b32 s85, s11, 8
; GFX9-NEXT:    s_lshr_b32 s84, s10, 16
; GFX9-NEXT:    s_lshr_b32 s86, s10, 8
; GFX9-NEXT:    s_lshr_b32 s87, s13, 24
; GFX9-NEXT:    s_lshr_b32 s96, s13, 16
; GFX9-NEXT:    s_lshr_b32 s98, s13, 8
; GFX9-NEXT:    s_lshr_b32 s97, s12, 16
; GFX9-NEXT:    s_lshr_b32 s99, s12, 8
; GFX9-NEXT:    s_lshr_b32 s38, s15, 24
; GFX9-NEXT:    s_lshr_b32 s39, s15, 16
; GFX9-NEXT:    s_lshr_b32 s49, s15, 8
; GFX9-NEXT:    s_lshr_b32 s48, s14, 16
; GFX9-NEXT:    s_lshr_b32 s50, s14, 8
; GFX9-NEXT:    s_lshr_b32 s51, s41, 24
; GFX9-NEXT:    s_lshr_b32 s52, s41, 16
; GFX9-NEXT:    s_lshr_b32 s54, s41, 8
; GFX9-NEXT:    s_lshr_b32 s53, s40, 16
; GFX9-NEXT:    s_lshr_b32 s55, s40, 8
; GFX9-NEXT:    s_lshr_b32 s64, s43, 24
; GFX9-NEXT:    s_lshr_b32 s65, s43, 16
; GFX9-NEXT:    s_lshr_b32 s67, s43, 8
; GFX9-NEXT:    s_lshr_b32 s66, s42, 16
; GFX9-NEXT:    s_lshr_b32 s68, s42, 8
; GFX9-NEXT:    s_lshr_b32 s69, s45, 24
; GFX9-NEXT:    s_lshr_b32 s70, s45, 16
; GFX9-NEXT:    s_lshr_b32 s80, s45, 8
; GFX9-NEXT:    s_lshr_b32 s71, s44, 16
; GFX9-NEXT:    s_lshr_b32 s81, s44, 8
; GFX9-NEXT:    s_lshr_b64 s[46:47], s[4:5], 24
; GFX9-NEXT:    s_lshr_b64 s[56:57], s[28:29], 24
; GFX9-NEXT:    s_lshr_b64 s[58:59], s[26:27], 24
; GFX9-NEXT:    s_lshr_b64 s[60:61], s[24:25], 24
; GFX9-NEXT:    s_lshr_b64 s[62:63], s[22:23], 24
; GFX9-NEXT:    s_lshr_b64 s[72:73], s[20:21], 24
; GFX9-NEXT:    s_lshr_b64 s[74:75], s[18:19], 24
; GFX9-NEXT:    s_lshr_b64 s[76:77], s[16:17], 24
; GFX9-NEXT:    s_lshr_b64 s[78:79], s[6:7], 24
; GFX9-NEXT:    s_lshr_b64 s[88:89], s[8:9], 24
; GFX9-NEXT:    s_lshr_b64 s[90:91], s[10:11], 24
; GFX9-NEXT:    s_lshr_b64 s[92:93], s[12:13], 24
; GFX9-NEXT:    s_lshr_b64 s[94:95], s[14:15], 24
; GFX9-NEXT:    s_lshr_b64 s[30:31], s[40:41], 24
; GFX9-NEXT:    s_lshr_b64 s[34:35], s[42:43], 24
; GFX9-NEXT:    s_lshr_b64 s[36:37], s[44:45], 24
; GFX9-NEXT:    s_cbranch_execnz .LBB99_4
; GFX9-NEXT:  .LBB99_2: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v26, s5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, s4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[25:26]
; GFX9-NEXT:    v_pk_add_u16 v28, s29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, s28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[27:28]
; GFX9-NEXT:    v_pk_add_u16 v30, s27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, s26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[29:30]
; GFX9-NEXT:    v_pk_add_u16 v32, s25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v31, s24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[31:32]
; GFX9-NEXT:    v_pk_add_u16 v34, s23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v33, s22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[33:34]
; GFX9-NEXT:    v_pk_add_u16 v38, s19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v36, s21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v35, s20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[35:36]
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v38
; GFX9-NEXT:    v_pk_add_u16 v49, s17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v38
; GFX9-NEXT:    v_pk_add_u16 v2, s7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, s6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v49
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[1:2]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v49
; GFX9-NEXT:    v_pk_add_u16 v4, s9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v1
; GFX9-NEXT:    v_pk_add_u16 v3, s8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v4
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[3:4]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v4
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v4
; GFX9-NEXT:    v_pk_add_u16 v6, s11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, s10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v3
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[5:6]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v3
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v6
; GFX9-NEXT:    v_pk_add_u16 v8, s13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, s12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v6
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[7:8]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v6
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v5
; GFX9-NEXT:    v_pk_add_u16 v10, s15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, s14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v5
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[9:10]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v8
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v8
; GFX9-NEXT:    v_pk_add_u16 v12, s41, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, s40, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v8
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[11:12]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v7
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v7
; GFX9-NEXT:    v_pk_add_u16 v14, s43, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, s42, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v10
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[13:14]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v10
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v10
; GFX9-NEXT:    v_pk_add_u16 v20, s45, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, s44, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v9
; GFX9-NEXT:    v_lshrrev_b64 v[15:16], 24, v[19:20]
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v9
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v12
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v26
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v12
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v26
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v26
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v11
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v25
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v28
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v28
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v30
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v30
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v32
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 24, v20
; GFX9-NEXT:    v_pk_add_u16 v37, s18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v32
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v20
; GFX9-NEXT:    v_pk_add_u16 v48, s16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_lshrrev_b64 v[43:44], 24, v[37:38]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 24, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 24, v36
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v20
; GFX9-NEXT:    v_lshrrev_b64 v[44:45], 24, v[48:49]
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 16, v34
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 16, v36
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 16, v19
; GFX9-NEXT:    v_lshrrev_b32_e32 v51, 16, v25
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 8, v28
; GFX9-NEXT:    v_lshrrev_b32_e32 v47, 16, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v46, 8, v27
; GFX9-NEXT:    v_lshrrev_b32_e32 v56, 8, v30
; GFX9-NEXT:    v_lshrrev_b32_e32 v58, 16, v29
; GFX9-NEXT:    v_lshrrev_b32_e32 v57, 8, v29
; GFX9-NEXT:    v_lshrrev_b32_e32 v59, 8, v32
; GFX9-NEXT:    v_lshrrev_b32_e32 v60, 16, v31
; GFX9-NEXT:    v_lshrrev_b32_e32 v53, 8, v31
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v23, 8, v34
; GFX9-NEXT:    v_lshrrev_b32_e32 v61, 16, v33
; GFX9-NEXT:    v_lshrrev_b32_e32 v15, 8, v33
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 8, v36
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v35
; GFX9-NEXT:    v_lshrrev_b32_e32 v21, 8, v35
; GFX9-NEXT:    v_lshrrev_b32_e32 v24, 8, v38
; GFX9-NEXT:    v_lshrrev_b32_e32 v54, 16, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v52, 8, v37
; GFX9-NEXT:    v_lshrrev_b32_e32 v50, 8, v49
; GFX9-NEXT:    v_lshrrev_b32_e32 v39, 16, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v55, 8, v48
; GFX9-NEXT:    v_lshrrev_b32_e32 v42, 24, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v40, 16, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v41, 8, v2
; GFX9-NEXT:    v_lshrrev_b32_e32 v45, 8, v1
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_lshrrev_b32_e32 v22, 8, v19
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    s_branch .LBB99_5
; GFX9-NEXT:  .LBB99_3:
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr81
; GFX9-NEXT:    ; implicit-def: $sgpr71
; GFX9-NEXT:    ; implicit-def: $sgpr80
; GFX9-NEXT:    ; implicit-def: $sgpr70
; GFX9-NEXT:    ; implicit-def: $sgpr69
; GFX9-NEXT:    ; implicit-def: $sgpr68
; GFX9-NEXT:    ; implicit-def: $sgpr66
; GFX9-NEXT:    ; implicit-def: $sgpr67
; GFX9-NEXT:    ; implicit-def: $sgpr65
; GFX9-NEXT:    ; implicit-def: $sgpr64
; GFX9-NEXT:    ; implicit-def: $sgpr55
; GFX9-NEXT:    ; implicit-def: $sgpr53
; GFX9-NEXT:    ; implicit-def: $sgpr54
; GFX9-NEXT:    ; implicit-def: $sgpr52
; GFX9-NEXT:    ; implicit-def: $sgpr51
; GFX9-NEXT:    ; implicit-def: $sgpr50
; GFX9-NEXT:    ; implicit-def: $sgpr48
; GFX9-NEXT:    ; implicit-def: $sgpr49
; GFX9-NEXT:    ; implicit-def: $sgpr39
; GFX9-NEXT:    ; implicit-def: $sgpr38
; GFX9-NEXT:    ; implicit-def: $sgpr99
; GFX9-NEXT:    ; implicit-def: $sgpr97
; GFX9-NEXT:    ; implicit-def: $sgpr98
; GFX9-NEXT:    ; implicit-def: $sgpr96
; GFX9-NEXT:    ; implicit-def: $sgpr87
; GFX9-NEXT:    ; implicit-def: $sgpr86
; GFX9-NEXT:    ; implicit-def: $sgpr84
; GFX9-NEXT:    ; implicit-def: $sgpr85
; GFX9-NEXT:    ; implicit-def: $sgpr83
; GFX9-NEXT:    ; implicit-def: $sgpr82
; GFX9-NEXT:    ; implicit-def: $sgpr76
; GFX9-NEXT:    ; implicit-def: $sgpr74
; GFX9-NEXT:    ; implicit-def: $sgpr72
; GFX9-NEXT:    ; implicit-def: $sgpr62
; GFX9-NEXT:    ; implicit-def: $sgpr60
; GFX9-NEXT:    ; implicit-def: $sgpr58
; GFX9-NEXT:    ; implicit-def: $sgpr56
; GFX9-NEXT:    ; implicit-def: $sgpr36
; GFX9-NEXT:    ; implicit-def: $sgpr34
; GFX9-NEXT:    ; implicit-def: $sgpr30
; GFX9-NEXT:    ; implicit-def: $sgpr94
; GFX9-NEXT:    ; implicit-def: $sgpr92
; GFX9-NEXT:    ; implicit-def: $sgpr90
; GFX9-NEXT:    ; implicit-def: $sgpr88
; GFX9-NEXT:    ; implicit-def: $sgpr78
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    ; kill: killed $sgpr46
; GFX9-NEXT:    ; implicit-def: $sgpr46
; GFX9-NEXT:    s_branch .LBB99_2
; GFX9-NEXT:  .LBB99_4:
; GFX9-NEXT:    v_mov_b32_e32 v15, s81
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s71
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s80
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s70
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s69
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s68
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s66
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s67
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s65
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s64
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s55
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s53
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s54
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s52
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s51
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s50
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s48
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s49
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s39
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s38
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s99
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s97
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s98
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s96
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s87
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s86
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s84
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s85
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s83
; GFX9-NEXT:    v_mov_b32_e32 v25, s4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s82
; GFX9-NEXT:    v_readlane_b32 s4, v62, 0
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 1
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 2
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 3
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 4
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 5
; GFX9-NEXT:    v_mov_b32_e32 v22, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 6
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 7
; GFX9-NEXT:    v_mov_b32_e32 v45, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 8
; GFX9-NEXT:    v_mov_b32_e32 v41, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 9
; GFX9-NEXT:    v_mov_b32_e32 v42, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 10
; GFX9-NEXT:    v_mov_b32_e32 v55, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 11
; GFX9-NEXT:    v_mov_b32_e32 v39, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 12
; GFX9-NEXT:    v_mov_b32_e32 v50, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 13
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 14
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 15
; GFX9-NEXT:    v_mov_b32_e32 v52, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 16
; GFX9-NEXT:    v_mov_b32_e32 v54, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 17
; GFX9-NEXT:    v_mov_b32_e32 v24, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 18
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 19
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 20
; GFX9-NEXT:    v_mov_b32_e32 v21, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 21
; GFX9-NEXT:    v_mov_b32_e32 v18, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 22
; GFX9-NEXT:    v_mov_b32_e32 v16, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 23
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 24
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 25
; GFX9-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v15, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 26
; GFX9-NEXT:    v_mov_b32_e32 v61, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 27
; GFX9-NEXT:    v_mov_b32_e32 v23, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 28
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 29
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 30
; GFX9-NEXT:    v_mov_b32_e32 v53, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 31
; GFX9-NEXT:    v_mov_b32_e32 v60, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 32
; GFX9-NEXT:    v_mov_b32_e32 v59, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 33
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 34
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 35
; GFX9-NEXT:    v_mov_b32_e32 v57, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 36
; GFX9-NEXT:    v_mov_b32_e32 v58, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 37
; GFX9-NEXT:    v_mov_b32_e32 v56, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 38
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 39
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 40
; GFX9-NEXT:    v_mov_b32_e32 v46, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 41
; GFX9-NEXT:    v_mov_b32_e32 v47, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 42
; GFX9-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v17, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 43
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 44
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 45
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 46
; GFX9-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v51, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 47
; GFX9-NEXT:    v_mov_b32_e32 v40, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 48
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s4
; GFX9-NEXT:    v_readlane_b32 s4, v62, 49
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s4
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s46
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s56
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s58
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s60
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s62
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s72
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s78
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v40, s88
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v19, s44
; GFX9-NEXT:    v_mov_b32_e32 v20, s45
; GFX9-NEXT:    v_mov_b32_e32 v13, s42
; GFX9-NEXT:    v_mov_b32_e32 v14, s43
; GFX9-NEXT:    v_mov_b32_e32 v11, s40
; GFX9-NEXT:    v_mov_b32_e32 v12, s41
; GFX9-NEXT:    v_mov_b32_e32 v9, s14
; GFX9-NEXT:    v_mov_b32_e32 v10, s15
; GFX9-NEXT:    v_mov_b32_e32 v7, s12
; GFX9-NEXT:    v_mov_b32_e32 v8, s13
; GFX9-NEXT:    v_mov_b32_e32 v5, s10
; GFX9-NEXT:    v_mov_b32_e32 v6, s11
; GFX9-NEXT:    v_mov_b32_e32 v3, s8
; GFX9-NEXT:    v_mov_b32_e32 v4, s9
; GFX9-NEXT:    v_mov_b32_e32 v1, s6
; GFX9-NEXT:    v_mov_b32_e32 v2, s7
; GFX9-NEXT:    v_mov_b32_e32 v48, s16
; GFX9-NEXT:    v_mov_b32_e32 v49, s17
; GFX9-NEXT:    v_mov_b32_e32 v37, s18
; GFX9-NEXT:    v_mov_b32_e32 v38, s19
; GFX9-NEXT:    v_mov_b32_e32 v35, s20
; GFX9-NEXT:    v_mov_b32_e32 v36, s21
; GFX9-NEXT:    v_mov_b32_e32 v33, s22
; GFX9-NEXT:    v_mov_b32_e32 v34, s23
; GFX9-NEXT:    v_mov_b32_e32 v31, s24
; GFX9-NEXT:    v_mov_b32_e32 v32, s25
; GFX9-NEXT:    v_mov_b32_e32 v29, s26
; GFX9-NEXT:    v_mov_b32_e32 v30, s27
; GFX9-NEXT:    v_mov_b32_e32 v27, s28
; GFX9-NEXT:    v_mov_b32_e32 v28, s29
; GFX9-NEXT:    v_mov_b32_e32 v26, s5
; GFX9-NEXT:    v_mov_b32_e32 v43, s74
; GFX9-NEXT:    v_mov_b32_e32 v44, s76
; GFX9-NEXT:    v_mov_b32_e32 v40, v41
; GFX9-NEXT:    v_mov_b32_e32 v41, v45
; GFX9-NEXT:    v_mov_b32_e32 v45, v22
; GFX9-NEXT:    v_mov_b32_e32 v22, s90
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s92
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s94
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s30
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s34
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; GFX9-NEXT:    v_mov_b32_e32 v22, s36
; GFX9-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; GFX9-NEXT:  .LBB99_5: ; %end
; GFX9-NEXT:    s_mov_b32 s4, 0xc0c0004
; GFX9-NEXT:    v_perm_b32 v37, v37, v52, s4
; GFX9-NEXT:    v_perm_b32 v31, v31, v53, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v33, v33, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v21, v35, v21, s4
; GFX9-NEXT:    v_perm_b32 v16, v36, v16, s4
; GFX9-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v24, v38, v24, s4
; GFX9-NEXT:    v_perm_b32 v17, v28, v17, s4
; GFX9-NEXT:    v_perm_b32 v48, v48, v55, s4
; GFX9-NEXT:    v_perm_b32 v49, v49, v50, s4
; GFX9-NEXT:    v_perm_b32 v50, v54, v43, s4
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v39, v39, v44, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX9-NEXT:    v_or_b32_e32 v39, v48, v39
; GFX9-NEXT:    v_perm_b32 v23, v34, v23, s4
; GFX9-NEXT:    v_perm_b32 v32, v32, v59, s4
; GFX9-NEXT:    v_perm_b32 v29, v29, v57, s4
; GFX9-NEXT:    v_perm_b32 v30, v30, v56, s4
; GFX9-NEXT:    v_perm_b32 v27, v27, v46, s4
; GFX9-NEXT:    v_perm_b32 v1, v1, v45, s4
; GFX9-NEXT:    v_readlane_b32 s99, v63, 35
; GFX9-NEXT:    v_readlane_b32 s98, v63, 34
; GFX9-NEXT:    v_readlane_b32 s97, v63, 33
; GFX9-NEXT:    v_readlane_b32 s96, v63, 32
; GFX9-NEXT:    v_readlane_b32 s87, v63, 31
; GFX9-NEXT:    v_readlane_b32 s86, v63, 30
; GFX9-NEXT:    v_readlane_b32 s85, v63, 29
; GFX9-NEXT:    v_readlane_b32 s84, v63, 28
; GFX9-NEXT:    v_readlane_b32 s83, v63, 27
; GFX9-NEXT:    v_readlane_b32 s82, v63, 26
; GFX9-NEXT:    v_readlane_b32 s81, v63, 25
; GFX9-NEXT:    v_readlane_b32 s80, v63, 24
; GFX9-NEXT:    v_readlane_b32 s71, v63, 23
; GFX9-NEXT:    v_readlane_b32 s70, v63, 22
; GFX9-NEXT:    v_readlane_b32 s69, v63, 21
; GFX9-NEXT:    v_readlane_b32 s68, v63, 20
; GFX9-NEXT:    v_readlane_b32 s67, v63, 19
; GFX9-NEXT:    v_readlane_b32 s66, v63, 18
; GFX9-NEXT:    v_readlane_b32 s65, v63, 17
; GFX9-NEXT:    v_readlane_b32 s64, v63, 16
; GFX9-NEXT:    v_readlane_b32 s55, v63, 15
; GFX9-NEXT:    v_readlane_b32 s54, v63, 14
; GFX9-NEXT:    v_readlane_b32 s53, v63, 13
; GFX9-NEXT:    v_readlane_b32 s52, v63, 12
; GFX9-NEXT:    v_readlane_b32 s51, v63, 11
; GFX9-NEXT:    v_readlane_b32 s50, v63, 10
; GFX9-NEXT:    v_readlane_b32 s49, v63, 9
; GFX9-NEXT:    v_readlane_b32 s48, v63, 8
; GFX9-NEXT:    v_readlane_b32 s39, v63, 7
; GFX9-NEXT:    v_readlane_b32 s38, v63, 6
; GFX9-NEXT:    v_readlane_b32 s37, v63, 5
; GFX9-NEXT:    v_readlane_b32 s36, v63, 4
; GFX9-NEXT:    v_readlane_b32 s35, v63, 3
; GFX9-NEXT:    v_readlane_b32 s34, v63, 2
; GFX9-NEXT:    v_readlane_b32 s31, v63, 1
; GFX9-NEXT:    v_readlane_b32 s30, v63, 0
; GFX9-NEXT:    s_waitcnt vmcnt(2)
; GFX9-NEXT:    v_perm_b32 v36, v58, v52, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v22, v25, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v35, v61, v35, s4
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v18, v18, v54, s4
; GFX9-NEXT:    buffer_load_dword v54, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v55, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_or_b32_e32 v18, v21, v18
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v38, v47, v52, s4
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v53, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(4)
; GFX9-NEXT:    v_perm_b32 v25, v26, v15, s4
; GFX9-NEXT:    s_waitcnt vmcnt(3)
; GFX9-NEXT:    v_perm_b32 v34, v60, v54, s4
; GFX9-NEXT:    s_waitcnt vmcnt(1)
; GFX9-NEXT:    v_perm_b32 v28, v51, v52, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v26, v15, v51, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v19, v19, v15, s4
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v20, v20, v15, s4
; GFX9-NEXT:    buffer_load_dword v51, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v15, v15, v51, s4
; GFX9-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; GFX9-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v39, v39, v48, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; GFX9-NEXT:    v_or_b32_e32 v39, v49, v39
; GFX9-NEXT:    buffer_store_dword v39, v0, s[0:3], 0 offen offset:4
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v50
; GFX9-NEXT:    v_or_b32_e32 v37, v37, v39
; GFX9-NEXT:    buffer_store_dword v37, v0, s[0:3], 0 offen offset:8
; GFX9-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v37, v37, v39, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; GFX9-NEXT:    v_or_b32_e32 v24, v24, v37
; GFX9-NEXT:    buffer_store_dword v24, v0, s[0:3], 0 offen offset:12
; GFX9-NEXT:    buffer_store_dword v18, v0, s[0:3], 0 offen offset:16
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v18, v18, v21, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_or_b32_e32 v16, v16, v18
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:20
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v35
; GFX9-NEXT:    v_or_b32_e32 v16, v33, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:24
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v23, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:28
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v34
; GFX9-NEXT:    v_or_b32_e32 v16, v31, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:32
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v32, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:36
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v36
; GFX9-NEXT:    v_or_b32_e32 v16, v29, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:40
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v30, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:44
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v38
; GFX9-NEXT:    v_or_b32_e32 v16, v27, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:48
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v18, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v17, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:52
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v28
; GFX9-NEXT:    v_or_b32_e32 v16, v22, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:56
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v25, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:60
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v26
; GFX9-NEXT:    v_or_b32_e32 v16, v19, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:64
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v16, v17, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_or_b32_e32 v16, v20, v16
; GFX9-NEXT:    buffer_store_dword v16, v0, s[0:3], 0 offen offset:68
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v13, v16, s4
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v15
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:72
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v14, v14, v15, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX9-NEXT:    v_or_b32_e32 v13, v13, v14
; GFX9-NEXT:    buffer_store_dword v13, v0, s[0:3], 0 offen offset:76
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v11, v13, s4
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v13, v14, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v13
; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:80
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s4
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v12, v12, v13, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX9-NEXT:    v_or_b32_e32 v11, v11, v12
; GFX9-NEXT:    buffer_store_dword v11, v0, s[0:3], 0 offen offset:84
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v9, v11, s4
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v11, v12, v11, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v11
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:88
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v10, v10, v11, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX9-NEXT:    v_or_b32_e32 v9, v9, v10
; GFX9-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:92
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v7, v9, s4
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v9, v10, v9, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v9
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:96
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s4
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v8, v8, v9, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX9-NEXT:    v_or_b32_e32 v7, v7, v8
; GFX9-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:100
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v5, v7, s4
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v7, v8, v7, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v7
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:104
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s4
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v6, v6, v7, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX9-NEXT:    v_or_b32_e32 v5, v5, v6
; GFX9-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:108
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v3, v5, s4
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v5, v6, v5, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v5
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:112
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v4, v4, v5, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX9-NEXT:    v_or_b32_e32 v3, v3, v4
; GFX9-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:116
; GFX9-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v3
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:120
; GFX9-NEXT:    v_perm_b32 v1, v2, v41, s4
; GFX9-NEXT:    v_perm_b32 v2, v40, v42, s4
; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_or_b32_e32 v1, v1, v2
; GFX9-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen offset:124
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    s_or_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v128i8_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v74, s32 offset:72
; GFX11-NEXT:    scratch_store_b32 off, v75, s32 offset:76
; GFX11-NEXT:    scratch_store_b32 off, v76, s32 offset:80
; GFX11-NEXT:    scratch_store_b32 off, v77, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v74, s30, 0
; GFX11-NEXT:    v_writelane_b32 v75, s96, 0
; GFX11-NEXT:    v_readfirstlane_b32 s42, v15
; GFX11-NEXT:    v_readfirstlane_b32 s5, v14
; GFX11-NEXT:    v_readfirstlane_b32 s4, v13
; GFX11-NEXT:    v_writelane_b32 v74, s31, 1
; GFX11-NEXT:    v_writelane_b32 v75, s97, 1
; GFX11-NEXT:    v_readfirstlane_b32 s7, v12
; GFX11-NEXT:    v_readfirstlane_b32 s6, v11
; GFX11-NEXT:    v_readfirstlane_b32 s9, v10
; GFX11-NEXT:    v_writelane_b32 v74, s34, 2
; GFX11-NEXT:    v_writelane_b32 v75, s98, 2
; GFX11-NEXT:    v_readfirstlane_b32 s8, v9
; GFX11-NEXT:    v_readfirstlane_b32 s11, v8
; GFX11-NEXT:    v_readfirstlane_b32 s10, v7
; GFX11-NEXT:    v_writelane_b32 v74, s35, 3
; GFX11-NEXT:    v_writelane_b32 v75, s99, 3
; GFX11-NEXT:    v_readfirstlane_b32 s13, v6
; GFX11-NEXT:    v_readfirstlane_b32 s12, v5
; GFX11-NEXT:    v_readfirstlane_b32 s15, v4
; GFX11-NEXT:    v_writelane_b32 v74, s36, 4
; GFX11-NEXT:    v_writelane_b32 v75, s100, 4
; GFX11-NEXT:    v_readfirstlane_b32 s14, v3
; GFX11-NEXT:    v_readfirstlane_b32 s41, v2
; GFX11-NEXT:    v_readfirstlane_b32 s40, v1
; GFX11-NEXT:    v_writelane_b32 v74, s37, 5
; GFX11-NEXT:    v_writelane_b32 v75, s101, 5
; GFX11-NEXT:    s_cmp_lg_u32 s42, 0
; GFX11-NEXT:    s_mov_b32 vcc_lo, 0
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Spill
; GFX11-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-NEXT:    scratch_store_b32 off, v72, s32 offset:4
; GFX11-NEXT:    scratch_store_b32 off, v73, s32
; GFX11-NEXT:    v_writelane_b32 v74, s38, 6
; GFX11-NEXT:    v_writelane_b32 v75, s102, 6
; GFX11-NEXT:    ; implicit-def: $vgpr76 : SGPR spill to VGPR lane
; GFX11-NEXT:    ; implicit-def: $vgpr77 : SGPR spill to VGPR lane
; GFX11-NEXT:    v_writelane_b32 v74, s39, 7
; GFX11-NEXT:    v_writelane_b32 v75, s103, 7
; GFX11-NEXT:    v_writelane_b32 v74, s48, 8
; GFX11-NEXT:    v_writelane_b32 v75, s104, 8
; GFX11-NEXT:    v_writelane_b32 v74, s49, 9
; GFX11-NEXT:    v_writelane_b32 v74, s50, 10
; GFX11-NEXT:    v_writelane_b32 v74, s51, 11
; GFX11-NEXT:    v_writelane_b32 v74, s52, 12
; GFX11-NEXT:    v_writelane_b32 v74, s53, 13
; GFX11-NEXT:    v_writelane_b32 v74, s54, 14
; GFX11-NEXT:    v_writelane_b32 v74, s55, 15
; GFX11-NEXT:    v_writelane_b32 v74, s64, 16
; GFX11-NEXT:    v_writelane_b32 v74, s65, 17
; GFX11-NEXT:    v_writelane_b32 v74, s66, 18
; GFX11-NEXT:    v_writelane_b32 v74, s67, 19
; GFX11-NEXT:    v_writelane_b32 v74, s68, 20
; GFX11-NEXT:    v_writelane_b32 v74, s69, 21
; GFX11-NEXT:    v_writelane_b32 v74, s70, 22
; GFX11-NEXT:    v_writelane_b32 v74, s71, 23
; GFX11-NEXT:    v_writelane_b32 v74, s80, 24
; GFX11-NEXT:    v_writelane_b32 v74, s81, 25
; GFX11-NEXT:    v_writelane_b32 v74, s82, 26
; GFX11-NEXT:    v_writelane_b32 v74, s83, 27
; GFX11-NEXT:    v_writelane_b32 v74, s84, 28
; GFX11-NEXT:    v_writelane_b32 v74, s85, 29
; GFX11-NEXT:    v_writelane_b32 v74, s86, 30
; GFX11-NEXT:    v_writelane_b32 v74, s87, 31
; GFX11-NEXT:    s_cbranch_scc0 .LBB99_3
; GFX11-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-NEXT:    s_lshr_b32 s42, s27, 24
; GFX11-NEXT:    s_lshr_b32 vcc_hi, s5, 24
; GFX11-NEXT:    v_writelane_b32 v77, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s27, 16
; GFX11-NEXT:    s_lshr_b32 s34, s5, 16
; GFX11-NEXT:    s_lshr_b32 s36, s5, 8
; GFX11-NEXT:    s_lshr_b32 s35, s4, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s27, 8
; GFX11-NEXT:    s_lshr_b32 s37, s4, 8
; GFX11-NEXT:    s_lshr_b32 s38, s7, 24
; GFX11-NEXT:    s_lshr_b32 s39, s7, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s26, 16
; GFX11-NEXT:    s_lshr_b32 s49, s7, 8
; GFX11-NEXT:    s_lshr_b32 s48, s6, 16
; GFX11-NEXT:    s_lshr_b32 s50, s6, 8
; GFX11-NEXT:    v_writelane_b32 v77, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s26, 8
; GFX11-NEXT:    s_lshr_b32 s51, s9, 24
; GFX11-NEXT:    s_lshr_b32 s52, s9, 16
; GFX11-NEXT:    s_lshr_b32 s54, s9, 8
; GFX11-NEXT:    v_writelane_b32 v77, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s25, 24
; GFX11-NEXT:    s_lshr_b32 s53, s8, 16
; GFX11-NEXT:    s_lshr_b32 s55, s8, 8
; GFX11-NEXT:    s_lshr_b32 s64, s11, 24
; GFX11-NEXT:    v_writelane_b32 v77, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s25, 16
; GFX11-NEXT:    s_lshr_b32 s65, s11, 16
; GFX11-NEXT:    s_lshr_b32 s67, s11, 8
; GFX11-NEXT:    s_lshr_b32 s66, s10, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s25, 8
; GFX11-NEXT:    s_lshr_b32 s68, s10, 8
; GFX11-NEXT:    s_lshr_b32 s69, s13, 24
; GFX11-NEXT:    s_lshr_b32 s70, s13, 16
; GFX11-NEXT:    v_writelane_b32 v77, s42, 0
; GFX11-NEXT:    s_lshr_b32 s42, s24, 16
; GFX11-NEXT:    s_lshr_b32 s80, s13, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 31
; GFX11-NEXT:    s_lshr_b32 s42, s24, 8
; GFX11-NEXT:    s_lshr_b32 s71, s12, 16
; GFX11-NEXT:    s_lshr_b32 s81, s12, 8
; GFX11-NEXT:    s_lshr_b32 s82, s15, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 30
; GFX11-NEXT:    s_lshr_b32 s42, s23, 24
; GFX11-NEXT:    s_lshr_b32 s83, s15, 16
; GFX11-NEXT:    s_lshr_b32 s85, s15, 8
; GFX11-NEXT:    s_lshr_b32 s84, s14, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 29
; GFX11-NEXT:    s_lshr_b32 s42, s23, 16
; GFX11-NEXT:    s_lshr_b32 s86, s14, 8
; GFX11-NEXT:    s_lshr_b32 s87, s41, 24
; GFX11-NEXT:    s_lshr_b32 s96, s41, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 28
; GFX11-NEXT:    s_lshr_b32 s42, s23, 8
; GFX11-NEXT:    s_lshr_b32 s98, s41, 8
; GFX11-NEXT:    s_lshr_b32 s97, s40, 16
; GFX11-NEXT:    s_lshr_b32 s99, s40, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 27
; GFX11-NEXT:    s_lshr_b32 s42, s22, 16
; GFX11-NEXT:    s_lshr_b32 s100, s29, 24
; GFX11-NEXT:    s_lshr_b32 s101, s29, 16
; GFX11-NEXT:    s_lshr_b32 s103, s29, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 26
; GFX11-NEXT:    s_lshr_b32 s42, s22, 8
; GFX11-NEXT:    s_lshr_b32 s102, s28, 16
; GFX11-NEXT:    s_lshr_b32 s104, s28, 8
; GFX11-NEXT:    s_lshr_b64 s[62:63], s[26:27], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 25
; GFX11-NEXT:    s_lshr_b32 s42, s21, 24
; GFX11-NEXT:    s_lshr_b64 s[72:73], s[24:25], 24
; GFX11-NEXT:    s_lshr_b64 s[60:61], s[22:23], 24
; GFX11-NEXT:    s_lshr_b64 s[58:59], s[20:21], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 24
; GFX11-NEXT:    s_lshr_b32 s42, s21, 16
; GFX11-NEXT:    s_lshr_b64 s[56:57], s[18:19], 24
; GFX11-NEXT:    s_lshr_b64 s[46:47], s[16:17], 24
; GFX11-NEXT:    s_lshr_b64 s[44:45], s[2:3], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 23
; GFX11-NEXT:    s_lshr_b32 s42, s21, 8
; GFX11-NEXT:    s_lshr_b64 s[74:75], s[4:5], 24
; GFX11-NEXT:    s_lshr_b64 s[76:77], s[6:7], 24
; GFX11-NEXT:    s_lshr_b64 s[78:79], s[8:9], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 22
; GFX11-NEXT:    s_lshr_b32 s42, s20, 16
; GFX11-NEXT:    s_lshr_b64 s[88:89], s[10:11], 24
; GFX11-NEXT:    s_lshr_b64 s[90:91], s[12:13], 24
; GFX11-NEXT:    s_lshr_b64 s[92:93], s[14:15], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 21
; GFX11-NEXT:    s_lshr_b32 s42, s20, 8
; GFX11-NEXT:    s_lshr_b64 s[94:95], s[40:41], 24
; GFX11-NEXT:    s_lshr_b64 s[30:31], s[28:29], 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 20
; GFX11-NEXT:    s_lshr_b32 s42, s19, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 19
; GFX11-NEXT:    s_lshr_b32 s42, s19, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 18
; GFX11-NEXT:    s_lshr_b32 s42, s19, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 17
; GFX11-NEXT:    s_lshr_b32 s42, s18, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 16
; GFX11-NEXT:    s_lshr_b32 s42, s18, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 15
; GFX11-NEXT:    s_lshr_b32 s42, s17, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 14
; GFX11-NEXT:    s_lshr_b32 s42, s17, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 13
; GFX11-NEXT:    s_lshr_b32 s42, s17, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 12
; GFX11-NEXT:    s_lshr_b32 s42, s16, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 11
; GFX11-NEXT:    s_lshr_b32 s42, s16, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 10
; GFX11-NEXT:    s_lshr_b32 s42, s3, 24
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 9
; GFX11-NEXT:    s_lshr_b32 s42, s3, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 8
; GFX11-NEXT:    s_lshr_b32 s42, s3, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 7
; GFX11-NEXT:    s_lshr_b32 s42, s2, 16
; GFX11-NEXT:    v_writelane_b32 v76, s42, 6
; GFX11-NEXT:    s_lshr_b32 s42, s2, 8
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 5
; GFX11-NEXT:    s_lshr_b32 s42, s1, 24
; GFX11-NEXT:    v_writelane_b32 v76, s42, 4
; GFX11-NEXT:    s_lshr_b32 s42, s1, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 3
; GFX11-NEXT:    s_lshr_b32 s42, s1, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 2
; GFX11-NEXT:    s_lshr_b32 s42, s0, 16
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
; GFX11-NEXT:    v_writelane_b32 v76, s42, 1
; GFX11-NEXT:    s_lshr_b32 s42, s0, 8
; GFX11-NEXT:    v_writelane_b32 v76, s42, 0
; GFX11-NEXT:    s_lshr_b64 s[42:43], s[0:1], 24
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, vcc_lo
; GFX11-NEXT:    s_cbranch_vccnz .LBB99_4
; GFX11-NEXT:  .LBB99_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v28, s19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, s27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, s26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, s18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, s17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, s16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, s25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, s24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, s41, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, s40, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, s13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, s12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, s9, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, s8, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v34, s3, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v33, s2, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, s23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, s22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, s29, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v15, s28, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, s15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, s14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, s11, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, s10, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, s7, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, s6, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, s5, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, s4, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v36, s1, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v35, s0, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, s21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, s20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_lshrrev_b64 v[51:52], 24, v[17:18]
; GFX11-NEXT:    v_lshrrev_b64 v[64:65], 24, v[27:28]
; GFX11-NEXT:    v_lshrrev_b64 v[52:53], 24, v[19:20]
; GFX11-NEXT:    v_lshrrev_b64 v[65:66], 24, v[29:30]
; GFX11-NEXT:    v_lshrrev_b64 v[53:54], 24, v[21:22]
; GFX11-NEXT:    v_lshrrev_b64 v[69:70], 24, v[33:34]
; GFX11-NEXT:    v_lshrrev_b64 v[37:38], 24, v[5:6]
; GFX11-NEXT:    v_lshrrev_b64 v[48:49], 24, v[9:10]
; GFX11-NEXT:    v_lshrrev_b64 v[66:67], 24, v[13:14]
; GFX11-NEXT:    v_lshrrev_b64 v[54:55], 24, v[23:24]
; GFX11-NEXT:    v_lshrrev_b64 v[70:71], 24, v[35:36]
; GFX11-NEXT:    v_lshrrev_b64 v[25:26], 24, v[1:2]
; GFX11-NEXT:    v_lshrrev_b64 v[31:32], 24, v[3:4]
; GFX11-NEXT:    v_lshrrev_b64 v[38:39], 24, v[7:8]
; GFX11-NEXT:    v_lshrrev_b64 v[49:50], 24, v[11:12]
; GFX11-NEXT:    v_lshrrev_b64 v[67:68], 24, v[15:16]
; GFX11-NEXT:    v_lshrrev_b32_e32 v148, 24, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v147, 16, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v146, 8, v18
; GFX11-NEXT:    v_lshrrev_b32_e32 v160, 16, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v150, 8, v17
; GFX11-NEXT:    v_lshrrev_b32_e32 v162, 24, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v161, 16, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v149, 8, v20
; GFX11-NEXT:    v_lshrrev_b32_e32 v163, 16, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v151, 8, v19
; GFX11-NEXT:    v_lshrrev_b32_e32 v166, 24, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v165, 16, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v164, 8, v22
; GFX11-NEXT:    v_lshrrev_b32_e32 v178, 16, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v176, 8, v21
; GFX11-NEXT:    v_lshrrev_b32_e32 v180, 24, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v179, 16, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v167, 8, v24
; GFX11-NEXT:    v_lshrrev_b32_e32 v181, 16, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v177, 8, v23
; GFX11-NEXT:    v_lshrrev_b32_e32 v42, 24, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v41, 16, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v183, 8, v28
; GFX11-NEXT:    v_lshrrev_b32_e32 v43, 16, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v182, 8, v27
; GFX11-NEXT:    v_lshrrev_b32_e32 v46, 24, v30
; GFX11-NEXT:    v_lshrrev_b32_e32 v44, 16, v30
; GFX11-NEXT:    v_lshrrev_b32_e32 v40, 8, v30
; GFX11-NEXT:    v_lshrrev_b32_e32 v47, 16, v29
; GFX11-NEXT:    v_lshrrev_b32_e32 v45, 8, v29
; GFX11-NEXT:    v_lshrrev_b32_e32 v59, 24, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v58, 16, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v57, 8, v34
; GFX11-NEXT:    v_lshrrev_b32_e32 v60, 16, v33
; GFX11-NEXT:    v_lshrrev_b32_e32 v56, 8, v33
; GFX11-NEXT:    v_lshrrev_b32_e32 v72, 24, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v62, 16, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v61, 8, v36
; GFX11-NEXT:    v_lshrrev_b32_e32 v73, 16, v35
; GFX11-NEXT:    v_lshrrev_b32_e32 v63, 8, v35
; GFX11-NEXT:    v_lshrrev_b32_e32 v50, 24, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v39, 16, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v26, 8, v2
; GFX11-NEXT:    v_lshrrev_b32_e32 v71, 16, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v32, 8, v1
; GFX11-NEXT:    v_lshrrev_b32_e32 v81, 24, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v80, 16, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v55, 8, v4
; GFX11-NEXT:    v_lshrrev_b32_e32 v84, 16, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v82, 8, v3
; GFX11-NEXT:    v_lshrrev_b32_e32 v86, 24, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v85, 16, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v83, 8, v6
; GFX11-NEXT:    v_lshrrev_b32_e32 v97, 16, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v87, 8, v5
; GFX11-NEXT:    v_lshrrev_b32_e32 v100, 24, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v99, 16, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v96, 8, v8
; GFX11-NEXT:    v_lshrrev_b32_e32 v101, 16, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v98, 8, v7
; GFX11-NEXT:    v_lshrrev_b32_e32 v112, 24, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v103, 16, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v102, 8, v10
; GFX11-NEXT:    v_lshrrev_b32_e32 v115, 16, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v113, 8, v9
; GFX11-NEXT:    v_lshrrev_b32_e32 v118, 24, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v117, 16, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v114, 8, v12
; GFX11-NEXT:    v_lshrrev_b32_e32 v119, 16, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v116, 8, v11
; GFX11-NEXT:    v_lshrrev_b32_e32 v130, 24, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v129, 16, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v128, 8, v14
; GFX11-NEXT:    v_lshrrev_b32_e32 v133, 16, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v131, 8, v13
; GFX11-NEXT:    v_lshrrev_b32_e32 v144, 24, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v135, 16, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v132, 8, v16
; GFX11-NEXT:    v_lshrrev_b32_e32 v145, 16, v15
; GFX11-NEXT:    v_lshrrev_b32_e32 v134, 8, v15
; GFX11-NEXT:    s_branch .LBB99_5
; GFX11-NEXT:  .LBB99_3:
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr42
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr44
; GFX11-NEXT:    ; implicit-def: $sgpr46
; GFX11-NEXT:    ; implicit-def: $sgpr56
; GFX11-NEXT:    ; implicit-def: $sgpr58
; GFX11-NEXT:    ; implicit-def: $sgpr60
; GFX11-NEXT:    ; implicit-def: $sgpr104
; GFX11-NEXT:    ; implicit-def: $sgpr102
; GFX11-NEXT:    ; implicit-def: $sgpr103
; GFX11-NEXT:    ; implicit-def: $sgpr101
; GFX11-NEXT:    ; implicit-def: $sgpr100
; GFX11-NEXT:    ; implicit-def: $sgpr99
; GFX11-NEXT:    ; implicit-def: $sgpr97
; GFX11-NEXT:    ; implicit-def: $sgpr98
; GFX11-NEXT:    ; implicit-def: $sgpr96
; GFX11-NEXT:    ; implicit-def: $sgpr87
; GFX11-NEXT:    ; implicit-def: $sgpr86
; GFX11-NEXT:    ; implicit-def: $sgpr84
; GFX11-NEXT:    ; implicit-def: $sgpr85
; GFX11-NEXT:    ; implicit-def: $sgpr83
; GFX11-NEXT:    ; implicit-def: $sgpr82
; GFX11-NEXT:    ; implicit-def: $sgpr81
; GFX11-NEXT:    ; implicit-def: $sgpr71
; GFX11-NEXT:    ; implicit-def: $sgpr80
; GFX11-NEXT:    ; implicit-def: $sgpr70
; GFX11-NEXT:    ; implicit-def: $sgpr69
; GFX11-NEXT:    ; implicit-def: $sgpr68
; GFX11-NEXT:    ; implicit-def: $sgpr66
; GFX11-NEXT:    ; implicit-def: $sgpr67
; GFX11-NEXT:    ; implicit-def: $sgpr65
; GFX11-NEXT:    ; implicit-def: $sgpr64
; GFX11-NEXT:    ; implicit-def: $sgpr55
; GFX11-NEXT:    ; implicit-def: $sgpr53
; GFX11-NEXT:    ; implicit-def: $sgpr54
; GFX11-NEXT:    ; implicit-def: $sgpr52
; GFX11-NEXT:    ; implicit-def: $sgpr51
; GFX11-NEXT:    ; implicit-def: $sgpr50
; GFX11-NEXT:    ; implicit-def: $sgpr48
; GFX11-NEXT:    ; implicit-def: $sgpr49
; GFX11-NEXT:    ; implicit-def: $sgpr39
; GFX11-NEXT:    ; implicit-def: $sgpr38
; GFX11-NEXT:    ; implicit-def: $sgpr37
; GFX11-NEXT:    ; implicit-def: $sgpr35
; GFX11-NEXT:    ; implicit-def: $sgpr36
; GFX11-NEXT:    ; implicit-def: $sgpr34
; GFX11-NEXT:    ; implicit-def: $vcc_hi
; GFX11-NEXT:    ; implicit-def: $sgpr72
; GFX11-NEXT:    ; implicit-def: $sgpr62
; GFX11-NEXT:    ; implicit-def: $sgpr30
; GFX11-NEXT:    ; implicit-def: $sgpr94
; GFX11-NEXT:    ; implicit-def: $sgpr92
; GFX11-NEXT:    ; implicit-def: $sgpr90
; GFX11-NEXT:    ; implicit-def: $sgpr88
; GFX11-NEXT:    ; implicit-def: $sgpr78
; GFX11-NEXT:    ; implicit-def: $sgpr76
; GFX11-NEXT:    ; implicit-def: $sgpr74
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    ; implicit-def: $sgpr43
; GFX11-NEXT:    ; kill: killed $sgpr43
; GFX11-NEXT:    s_branch .LBB99_2
; GFX11-NEXT:  .LBB99_4:
; GFX11-NEXT:    v_dual_mov_b32 v35, s0 :: v_dual_mov_b32 v36, s1
; GFX11-NEXT:    v_readlane_b32 s0, v76, 0
; GFX11-NEXT:    v_dual_mov_b32 v15, s28 :: v_dual_mov_b32 v16, s29
; GFX11-NEXT:    v_dual_mov_b32 v13, s40 :: v_dual_mov_b32 v14, s41
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v63, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 1
; GFX11-NEXT:    v_dual_mov_b32 v11, s14 :: v_dual_mov_b32 v12, s15
; GFX11-NEXT:    v_dual_mov_b32 v9, s12 :: v_dual_mov_b32 v10, s13
; GFX11-NEXT:    v_mov_b32_e32 v73, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 2
; GFX11-NEXT:    v_dual_mov_b32 v7, s10 :: v_dual_mov_b32 v8, s11
; GFX11-NEXT:    v_dual_mov_b32 v5, s8 :: v_dual_mov_b32 v6, s9
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v61, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 3
; GFX11-NEXT:    v_dual_mov_b32 v3, s6 :: v_dual_mov_b32 v4, s7
; GFX11-NEXT:    v_dual_mov_b32 v1, s4 :: v_dual_mov_b32 v2, s5
; GFX11-NEXT:    v_mov_b32_e32 v62, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 4
; GFX11-NEXT:    v_dual_mov_b32 v33, s2 :: v_dual_mov_b32 v34, s3
; GFX11-NEXT:    v_dual_mov_b32 v29, s16 :: v_dual_mov_b32 v30, s17
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v72, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 5
; GFX11-NEXT:    v_dual_mov_b32 v27, s18 :: v_dual_mov_b32 v28, s19
; GFX11-NEXT:    v_dual_mov_b32 v23, s20 :: v_dual_mov_b32 v24, s21
; GFX11-NEXT:    v_mov_b32_e32 v56, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 6
; GFX11-NEXT:    v_dual_mov_b32 v21, s22 :: v_dual_mov_b32 v22, s23
; GFX11-NEXT:    v_dual_mov_b32 v19, s24 :: v_dual_mov_b32 v20, s25
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v60, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 7
; GFX11-NEXT:    v_dual_mov_b32 v17, s26 :: v_dual_mov_b32 v18, s27
; GFX11-NEXT:    v_dual_mov_b32 v134, s104 :: v_dual_mov_b32 v145, s102
; GFX11-NEXT:    v_mov_b32_e32 v57, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 8
; GFX11-NEXT:    v_dual_mov_b32 v132, s103 :: v_dual_mov_b32 v135, s101
; GFX11-NEXT:    v_dual_mov_b32 v144, s100 :: v_dual_mov_b32 v131, s99
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v58, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 9
; GFX11-NEXT:    v_dual_mov_b32 v133, s97 :: v_dual_mov_b32 v128, s98
; GFX11-NEXT:    v_dual_mov_b32 v129, s96 :: v_dual_mov_b32 v130, s87
; GFX11-NEXT:    v_mov_b32_e32 v59, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 10
; GFX11-NEXT:    v_dual_mov_b32 v116, s86 :: v_dual_mov_b32 v119, s84
; GFX11-NEXT:    v_dual_mov_b32 v114, s85 :: v_dual_mov_b32 v117, s83
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v45, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 11
; GFX11-NEXT:    v_dual_mov_b32 v118, s82 :: v_dual_mov_b32 v113, s81
; GFX11-NEXT:    v_dual_mov_b32 v115, s71 :: v_dual_mov_b32 v102, s80
; GFX11-NEXT:    v_mov_b32_e32 v47, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 12
; GFX11-NEXT:    v_dual_mov_b32 v103, s70 :: v_dual_mov_b32 v112, s69
; GFX11-NEXT:    v_dual_mov_b32 v98, s68 :: v_dual_mov_b32 v101, s66
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v40, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 13
; GFX11-NEXT:    v_dual_mov_b32 v96, s67 :: v_dual_mov_b32 v99, s65
; GFX11-NEXT:    v_dual_mov_b32 v100, s64 :: v_dual_mov_b32 v87, s55
; GFX11-NEXT:    v_mov_b32_e32 v44, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 14
; GFX11-NEXT:    v_dual_mov_b32 v97, s53 :: v_dual_mov_b32 v86, s51
; GFX11-NEXT:    v_dual_mov_b32 v83, s54 :: v_dual_mov_b32 v82, s50
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v46, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 15
; GFX11-NEXT:    v_dual_mov_b32 v85, s52 :: v_dual_mov_b32 v84, s48
; GFX11-NEXT:    v_dual_mov_b32 v55, s49 :: v_dual_mov_b32 v80, s39
; GFX11-NEXT:    v_mov_b32_e32 v182, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 16
; GFX11-NEXT:    v_dual_mov_b32 v81, s38 :: v_dual_mov_b32 v32, s37
; GFX11-NEXT:    v_dual_mov_b32 v71, s35 :: v_dual_mov_b32 v26, s36
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v43, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 17
; GFX11-NEXT:    v_dual_mov_b32 v39, s34 :: v_dual_mov_b32 v50, vcc_hi
; GFX11-NEXT:    v_dual_mov_b32 v51, s62 :: v_dual_mov_b32 v52, s72
; GFX11-NEXT:    v_mov_b32_e32 v183, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 18
; GFX11-NEXT:    v_dual_mov_b32 v53, s60 :: v_dual_mov_b32 v54, s58
; GFX11-NEXT:    v_dual_mov_b32 v64, s56 :: v_dual_mov_b32 v65, s46
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v41, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 19
; GFX11-NEXT:    v_dual_mov_b32 v69, s44 :: v_dual_mov_b32 v70, s42
; GFX11-NEXT:    v_dual_mov_b32 v25, s74 :: v_dual_mov_b32 v38, s88
; GFX11-NEXT:    v_mov_b32_e32 v42, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 20
; GFX11-NEXT:    v_dual_mov_b32 v31, s76 :: v_dual_mov_b32 v48, s90
; GFX11-NEXT:    v_dual_mov_b32 v37, s78 :: v_dual_mov_b32 v66, s94
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-NEXT:    v_mov_b32_e32 v177, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 21
; GFX11-NEXT:    v_mov_b32_e32 v49, s92
; GFX11-NEXT:    v_mov_b32_e32 v67, s30
; GFX11-NEXT:    v_mov_b32_e32 v181, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 22
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v167, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 23
; GFX11-NEXT:    v_mov_b32_e32 v179, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 24
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v180, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 25
; GFX11-NEXT:    v_mov_b32_e32 v176, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 26
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v178, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 27
; GFX11-NEXT:    v_mov_b32_e32 v164, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 28
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v165, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 29
; GFX11-NEXT:    v_mov_b32_e32 v166, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 30
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v151, s0
; GFX11-NEXT:    v_readlane_b32 s0, v76, 31
; GFX11-NEXT:    v_mov_b32_e32 v163, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 0
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v149, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 1
; GFX11-NEXT:    v_mov_b32_e32 v161, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 2
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v162, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 3
; GFX11-NEXT:    v_mov_b32_e32 v150, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 4
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v160, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 5
; GFX11-NEXT:    v_mov_b32_e32 v146, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 6
; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-NEXT:    v_mov_b32_e32 v147, s0
; GFX11-NEXT:    v_readlane_b32 s0, v77, 7
; GFX11-NEXT:    v_mov_b32_e32 v148, s0
; GFX11-NEXT:  .LBB99_5: ; %end
; GFX11-NEXT:    v_perm_b32 v68, v73, v70, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v70, v62, v72, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v69, v60, v69, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v36, v36, v61, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v65, v47, v65, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v64, v43, v64, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v70, 16, v70
; GFX11-NEXT:    v_perm_b32 v35, v35, v63, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 16, v68
; GFX11-NEXT:    v_perm_b32 v56, v33, v56, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
; GFX11-NEXT:    v_perm_b32 v58, v58, v59, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v57, v34, v57, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v34, v36, v70
; GFX11-NEXT:    v_perm_b32 v29, v29, v45, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v65, 16, v65
; GFX11-NEXT:    v_perm_b32 v70, v27, v182, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX11-NEXT:    v_or_b32_e32 v33, v35, v68
; GFX11-NEXT:    v_or_b32_e32 v35, v56, v69
; GFX11-NEXT:    v_perm_b32 v68, v44, v46, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v69, v41, v42, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v58, 16, v58
; GFX11-NEXT:    v_or_b32_e32 v27, v29, v65
; GFX11-NEXT:    v_or_b32_e32 v29, v70, v64
; GFX11-NEXT:    v_perm_b32 v54, v181, v54, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v64, v179, v180, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v53, v178, v53, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v30, v30, v40, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v68, 16, v68
; GFX11-NEXT:    v_perm_b32 v182, v28, v183, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v69, 16, v69
; GFX11-NEXT:    v_or_b32_e32 v36, v57, v58
; GFX11-NEXT:    v_perm_b32 v23, v23, v177, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; GFX11-NEXT:    v_perm_b32 v24, v24, v167, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v64, 16, v64
; GFX11-NEXT:    v_perm_b32 v21, v21, v176, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; GFX11-NEXT:    v_or_b32_e32 v28, v30, v68
; GFX11-NEXT:    v_or_b32_e32 v30, v182, v69
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_store_b128 v0, v[33:36], off
; GFX11-NEXT:    scratch_store_b128 v0, v[27:30], off offset:16
; GFX11-NEXT:    v_or_b32_e32 v27, v23, v54
; GFX11-NEXT:    v_perm_b32 v23, v165, v166, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v28, v24, v64
; GFX11-NEXT:    v_or_b32_e32 v29, v21, v53
; GFX11-NEXT:    v_perm_b32 v21, v22, v164, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v22, v163, v52, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v161, v162, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v30, v160, v51, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v19, v19, v151, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v20, v20, v149, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v17, v17, v150, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v33, 16, v30
; GFX11-NEXT:    v_or_b32_e32 v30, v21, v23
; GFX11-NEXT:    v_or_b32_e32 v19, v19, v22
; GFX11-NEXT:    v_perm_b32 v22, v147, v148, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v20, v20, v24
; GFX11-NEXT:    v_or_b32_e32 v21, v17, v33
; GFX11-NEXT:    v_perm_b32 v17, v18, v146, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v18, v145, v67, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v133, v66, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v135, v144, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; GFX11-NEXT:    v_perm_b32 v15, v15, v134, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v13, v13, v131, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v16, v16, v132, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_or_b32_e32 v22, v17, v22
; GFX11-NEXT:    v_or_b32_e32 v15, v15, v18
; GFX11-NEXT:    v_perm_b32 v18, v129, v130, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v17, v13, v24
; GFX11-NEXT:    v_perm_b32 v24, v115, v48, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v16, v16, v23
; GFX11-NEXT:    v_perm_b32 v13, v14, v128, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v14, v119, v49, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v117, v118, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; GFX11-NEXT:    v_perm_b32 v9, v9, v113, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v11, v11, v116, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v12, v12, v114, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_or_b32_e32 v18, v13, v18
; GFX11-NEXT:    v_or_b32_e32 v13, v9, v24
; GFX11-NEXT:    v_perm_b32 v9, v10, v102, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v10, v101, v38, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v11, v11, v14
; GFX11-NEXT:    v_perm_b32 v14, v103, v112, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v12, v12, v23
; GFX11-NEXT:    v_perm_b32 v23, v99, v100, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v97, v37, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v7, v7, v98, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-NEXT:    v_perm_b32 v8, v8, v96, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v5, v5, v87, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_or_b32_e32 v7, v7, v10
; GFX11-NEXT:    v_perm_b32 v10, v85, v86, 0xc0c0004
; GFX11-NEXT:    v_or_b32_e32 v14, v9, v14
; GFX11-NEXT:    v_or_b32_e32 v8, v8, v23
; GFX11-NEXT:    v_or_b32_e32 v9, v5, v24
; GFX11-NEXT:    v_perm_b32 v5, v84, v31, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v23, v80, v81, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v24, v71, v25, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v25, v39, v50, 0xc0c0004
; GFX11-NEXT:    v_perm_b32 v6, v6, v83, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-NEXT:    v_perm_b32 v3, v3, v82, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; GFX11-NEXT:    v_perm_b32 v4, v4, v55, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; GFX11-NEXT:    v_perm_b32 v31, v1, v32, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; GFX11-NEXT:    v_perm_b32 v26, v2, v26, 0xc0c0004
; GFX11-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; GFX11-NEXT:    v_or_b32_e32 v10, v6, v10
; GFX11-NEXT:    v_or_b32_e32 v1, v3, v5
; GFX11-NEXT:    v_or_b32_e32 v2, v4, v23
; GFX11-NEXT:    v_or_b32_e32 v3, v31, v24
; GFX11-NEXT:    v_or_b32_e32 v4, v26, v25
; GFX11-NEXT:    s_clause 0x5
; GFX11-NEXT:    scratch_store_b128 v0, v[27:30], off offset:32
; GFX11-NEXT:    scratch_store_b128 v0, v[19:22], off offset:48
; GFX11-NEXT:    scratch_store_b128 v0, v[15:18], off offset:64
; GFX11-NEXT:    scratch_store_b128 v0, v[11:14], off offset:80
; GFX11-NEXT:    scratch_store_b128 v0, v[7:10], off offset:96
; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off offset:112
; GFX11-NEXT:    s_clause 0x11 ; 72-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v73, off, s32
; GFX11-NEXT:    scratch_load_b32 v72, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-NEXT:    v_readlane_b32 s104, v75, 8
; GFX11-NEXT:    v_readlane_b32 s103, v75, 7
; GFX11-NEXT:    v_readlane_b32 s102, v75, 6
; GFX11-NEXT:    v_readlane_b32 s101, v75, 5
; GFX11-NEXT:    v_readlane_b32 s100, v75, 4
; GFX11-NEXT:    v_readlane_b32 s99, v75, 3
; GFX11-NEXT:    v_readlane_b32 s98, v75, 2
; GFX11-NEXT:    v_readlane_b32 s97, v75, 1
; GFX11-NEXT:    v_readlane_b32 s96, v75, 0
; GFX11-NEXT:    v_readlane_b32 s87, v74, 31
; GFX11-NEXT:    v_readlane_b32 s86, v74, 30
; GFX11-NEXT:    v_readlane_b32 s85, v74, 29
; GFX11-NEXT:    v_readlane_b32 s84, v74, 28
; GFX11-NEXT:    v_readlane_b32 s83, v74, 27
; GFX11-NEXT:    v_readlane_b32 s82, v74, 26
; GFX11-NEXT:    v_readlane_b32 s81, v74, 25
; GFX11-NEXT:    v_readlane_b32 s80, v74, 24
; GFX11-NEXT:    v_readlane_b32 s71, v74, 23
; GFX11-NEXT:    v_readlane_b32 s70, v74, 22
; GFX11-NEXT:    v_readlane_b32 s69, v74, 21
; GFX11-NEXT:    v_readlane_b32 s68, v74, 20
; GFX11-NEXT:    v_readlane_b32 s67, v74, 19
; GFX11-NEXT:    v_readlane_b32 s66, v74, 18
; GFX11-NEXT:    v_readlane_b32 s65, v74, 17
; GFX11-NEXT:    v_readlane_b32 s64, v74, 16
; GFX11-NEXT:    v_readlane_b32 s55, v74, 15
; GFX11-NEXT:    v_readlane_b32 s54, v74, 14
; GFX11-NEXT:    v_readlane_b32 s53, v74, 13
; GFX11-NEXT:    v_readlane_b32 s52, v74, 12
; GFX11-NEXT:    v_readlane_b32 s51, v74, 11
; GFX11-NEXT:    v_readlane_b32 s50, v74, 10
; GFX11-NEXT:    v_readlane_b32 s49, v74, 9
; GFX11-NEXT:    v_readlane_b32 s48, v74, 8
; GFX11-NEXT:    v_readlane_b32 s39, v74, 7
; GFX11-NEXT:    v_readlane_b32 s38, v74, 6
; GFX11-NEXT:    v_readlane_b32 s37, v74, 5
; GFX11-NEXT:    v_readlane_b32 s36, v74, 4
; GFX11-NEXT:    v_readlane_b32 s35, v74, 3
; GFX11-NEXT:    v_readlane_b32 s34, v74, 2
; GFX11-NEXT:    v_readlane_b32 s31, v74, 1
; GFX11-NEXT:    v_readlane_b32 s30, v74, 0
; GFX11-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-NEXT:    s_clause 0x3 ; 16-byte Folded Reload
; GFX11-NEXT:    scratch_load_b32 v74, off, s32 offset:72
; GFX11-NEXT:    scratch_load_b32 v75, off, s32 offset:76
; GFX11-NEXT:    scratch_load_b32 v76, off, s32 offset:80
; GFX11-NEXT:    scratch_load_b32 v77, off, s32 offset:84
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <128 x i8>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <128 x i8>
  br label %end

end:
  %phi = phi <128 x i8> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <128 x i8> %phi
}

define <64 x half> @bitcast_v64bf16_to_v64f16(<64 x bfloat> %a, i32 %b) {
; SI-LABEL: bitcast_v64bf16_to_v64f16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:4
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v29
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v3
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v28
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v27
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v5
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v3
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v26
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v23
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v7
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v5
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v8
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v10
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v7
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v10
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v11
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v9
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v12
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v13
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v11
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v14
; SI-NEXT:    v_and_b32_e32 v46, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v15
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v13
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v16
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v15
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v45
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v19
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v17
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v42
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v19
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v55
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v52
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v25
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v53
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v25
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v49
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v50
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v38
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v39
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v35
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v34
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v24
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v37
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v57
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v36
; SI-NEXT:    v_mul_f32_e32 v45, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v2
; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v4
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v23
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v25
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v29
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v31
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v62
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v59
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v37
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v46
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v40
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v51
; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v33
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v32
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v6
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v63
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v61
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v60
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v58
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v47
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v44
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v41
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; kill: killed $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; kill: killed $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; kill: killed $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; kill: killed $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; kill: killed $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; kill: killed $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; kill: killed $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; kill: killed $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB100_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v30
; SI-NEXT:    v_alignbit_b32 v2, v2, v45, 16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v28
; SI-NEXT:    v_alignbit_b32 v2, v2, v7, 16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v26
; SI-NEXT:    v_alignbit_b32 v2, v2, v43, 16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v24
; SI-NEXT:    v_alignbit_b32 v1, v2, v1, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v22
; SI-NEXT:    v_alignbit_b32 v1, v1, v40, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v20
; SI-NEXT:    v_alignbit_b32 v1, v1, v54, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v18
; SI-NEXT:    v_alignbit_b32 v1, v1, v11, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v16
; SI-NEXT:    v_alignbit_b32 v1, v1, v50, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v14
; SI-NEXT:    v_alignbit_b32 v1, v1, v3, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v12
; SI-NEXT:    v_alignbit_b32 v1, v1, v48, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v10
; SI-NEXT:    v_alignbit_b32 v1, v1, v19, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v8
; SI-NEXT:    v_alignbit_b32 v1, v1, v9, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v13
; SI-NEXT:    v_alignbit_b32 v1, v1, v17, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v62
; SI-NEXT:    v_alignbit_b32 v1, v1, v5, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v59
; SI-NEXT:    v_alignbit_b32 v1, v1, v15, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v57
; SI-NEXT:    v_alignbit_b32 v0, v1, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v6, v2, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v30
; SI-NEXT:    v_alignbit_b32 v0, v6, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v4, v63, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v28
; SI-NEXT:    v_alignbit_b32 v0, v4, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v61, v60, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v26
; SI-NEXT:    v_alignbit_b32 v0, v61, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v58, v56, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v24
; SI-NEXT:    v_alignbit_b32 v0, v58, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v47, v46, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v22
; SI-NEXT:    v_alignbit_b32 v0, v47, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v44, v42, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v20
; SI-NEXT:    v_alignbit_b32 v0, v44, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v41, v55, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v18
; SI-NEXT:    v_alignbit_b32 v0, v41, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v53, v52, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v16
; SI-NEXT:    v_alignbit_b32 v0, v53, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v51, v49, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v14
; SI-NEXT:    v_alignbit_b32 v0, v51, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v39, v38, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v12
; SI-NEXT:    v_alignbit_b32 v0, v39, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v37, v21, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v10
; SI-NEXT:    v_alignbit_b32 v0, v37, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v36, v23, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v8
; SI-NEXT:    v_alignbit_b32 v0, v36, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v35, v25, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v13
; SI-NEXT:    v_alignbit_b32 v0, v35, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v34, v27, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v62
; SI-NEXT:    v_alignbit_b32 v0, v34, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v33, v29, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v59
; SI-NEXT:    v_alignbit_b32 v0, v33, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_alignbit_b32 v32, v31, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v57
; SI-NEXT:    v_alignbit_b32 v0, v32, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:  .LBB100_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB100_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v30
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v30
; SI-NEXT:    v_alignbit_b32 v2, v4, v2, 16
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v7
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v28
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
; SI-NEXT:    v_alignbit_b32 v4, v6, v4, 16
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v26
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
; SI-NEXT:    v_alignbit_b32 v6, v7, v6, 16
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v24
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
; SI-NEXT:    v_alignbit_b32 v6, v7, v6, 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v40
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v22
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v6
; SI-NEXT:    v_alignbit_b32 v7, v21, v7, 16
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v54
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v7
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v18
; SI-NEXT:    v_alignbit_b32 v11, v20, v11, 16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v16
; SI-NEXT:    v_alignbit_b32 v11, v20, v11, 16
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v14
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v11
; SI-NEXT:    v_alignbit_b32 v3, v14, v3, 16
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v48
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v12
; SI-NEXT:    v_alignbit_b32 v3, v14, v3, 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v19
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v10
; SI-NEXT:    v_alignbit_b32 v3, v14, v3, 16
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v9
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
; SI-NEXT:    v_alignbit_b32 v3, v9, v3, 16
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v13
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v9
; SI-NEXT:    v_alignbit_b32 v3, v13, v3, 16
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v5
; SI-NEXT:    v_alignbit_b32 v3, v13, v3, 16
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v59
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; SI-NEXT:    v_alignbit_b32 v3, v14, v3, 16
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v57
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v3
; SI-NEXT:    v_alignbit_b32 v0, v14, v0, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_alignbit_b32 v32, v31, v0, 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v0, v32, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_alignbit_b32 v33, v29, v3, 16
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v13
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v34, v27, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v35, v25, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v36, v23, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v37, v21, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v39, v38, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v51, v49, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_alignbit_b32 v53, v52, v13, 16
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v44, v42, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v47, v46, v15, 16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_alignbit_b32 v58, v56, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v58, v1, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v47, v15, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v44, v7, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_alignbit_b32 v61, v60, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_alignbit_b32 v41, v55, v14, 16
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v18
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v41, v14, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v53, v13, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v51, v11, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v39, v12, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v37, v10, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v36, v8, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v35, v9, 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v30
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v34, v5, 16
; SI-NEXT:    v_alignbit_b32 v16, v61, v16, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v3, 16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v6
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_alignbit_b32 v4, v63, v4, 16
; SI-NEXT:    v_alignbit_b32 v17, v4, v17, 16
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_alignbit_b32 v6, v2, v6, 16
; SI-NEXT:    v_alignbit_b32 v18, v6, v18, 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:  .LBB100_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v56
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v46
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v55
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v52
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v49
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v38
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_and_b32_e32 v32, 0xffff, v32
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_or_b32_e32 v31, v32, v31
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v60
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v63
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v61
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v58
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v47
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v44
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v41
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v53
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v51
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v37
; SI-NEXT:    v_or_b32_e32 v21, v22, v21
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v36
; SI-NEXT:    v_or_b32_e32 v23, v24, v23
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v35
; SI-NEXT:    v_or_b32_e32 v25, v26, v25
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v34
; SI-NEXT:    v_or_b32_e32 v27, v28, v27
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v33
; SI-NEXT:    v_or_b32_e32 v29, v30, v29
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_or_b32_e32 v30, v30, v33
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v64f16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB100_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v16, 16, 1
; VI-NEXT:    s_movk_i32 s6, 0x7fff
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v16
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v17
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v34, 16, v18
; VI-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; VI-NEXT:    v_bfe_u32 v35, v34, 16, 1
; VI-NEXT:    v_add_u32_e32 v35, vcc, v35, v34
; VI-NEXT:    v_add_u32_e32 v35, vcc, s6, v35
; VI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; VI-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; VI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; VI-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc
; VI-NEXT:    v_bfe_u32 v35, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v35, vcc, v35, v18
; VI-NEXT:    v_add_u32_e32 v35, vcc, s6, v35
; VI-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v35, v36, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
; VI-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; VI-NEXT:    v_bfe_u32 v36, v35, 16, 1
; VI-NEXT:    v_add_u32_e32 v36, vcc, v36, v35
; VI-NEXT:    v_add_u32_e32 v36, vcc, s6, v36
; VI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; VI-NEXT:    v_or_b32_e32 v37, 0x400000, v35
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; VI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; VI-NEXT:    v_cndmask_b32_e32 v35, v36, v37, vcc
; VI-NEXT:    v_bfe_u32 v36, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v36, vcc, v36, v19
; VI-NEXT:    v_add_u32_e32 v36, vcc, s6, v36
; VI-NEXT:    v_or_b32_e32 v37, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_cndmask_b32_e32 v19, v36, v37, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
; VI-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
; VI-NEXT:    v_bfe_u32 v37, v36, 16, 1
; VI-NEXT:    v_add_u32_e32 v37, vcc, v37, v36
; VI-NEXT:    v_add_u32_e32 v37, vcc, s6, v37
; VI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; VI-NEXT:    v_or_b32_e32 v38, 0x400000, v36
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; VI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; VI-NEXT:    v_cndmask_b32_e32 v36, v37, v38, vcc
; VI-NEXT:    v_bfe_u32 v37, v20, 16, 1
; VI-NEXT:    v_add_u32_e32 v37, vcc, v37, v20
; VI-NEXT:    v_add_u32_e32 v37, vcc, s6, v37
; VI-NEXT:    v_or_b32_e32 v38, 0x400000, v20
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; VI-NEXT:    v_cndmask_b32_e32 v20, v37, v38, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; VI-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
; VI-NEXT:    v_bfe_u32 v38, v37, 16, 1
; VI-NEXT:    v_add_u32_e32 v38, vcc, v38, v37
; VI-NEXT:    v_add_u32_e32 v38, vcc, s6, v38
; VI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; VI-NEXT:    v_or_b32_e32 v39, 0x400000, v37
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; VI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; VI-NEXT:    v_cndmask_b32_e32 v37, v38, v39, vcc
; VI-NEXT:    v_bfe_u32 v38, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v38, vcc, v38, v21
; VI-NEXT:    v_add_u32_e32 v38, vcc, s6, v38
; VI-NEXT:    v_or_b32_e32 v39, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_cndmask_b32_e32 v21, v38, v39, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v38, 16, v22
; VI-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; VI-NEXT:    v_bfe_u32 v39, v38, 16, 1
; VI-NEXT:    v_add_u32_e32 v39, vcc, v39, v38
; VI-NEXT:    v_add_u32_e32 v39, vcc, s6, v39
; VI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; VI-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; VI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; VI-NEXT:    v_cndmask_b32_e32 v38, v39, v48, vcc
; VI-NEXT:    v_bfe_u32 v39, v22, 16, 1
; VI-NEXT:    v_add_u32_e32 v39, vcc, v39, v22
; VI-NEXT:    v_add_u32_e32 v39, vcc, s6, v39
; VI-NEXT:    v_or_b32_e32 v48, 0x400000, v22
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; VI-NEXT:    v_cndmask_b32_e32 v22, v39, v48, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
; VI-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
; VI-NEXT:    v_bfe_u32 v48, v39, 16, 1
; VI-NEXT:    v_add_u32_e32 v48, vcc, v48, v39
; VI-NEXT:    v_add_u32_e32 v48, vcc, s6, v48
; VI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; VI-NEXT:    v_or_b32_e32 v49, 0x400000, v39
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; VI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; VI-NEXT:    v_cndmask_b32_e32 v39, v48, v49, vcc
; VI-NEXT:    v_bfe_u32 v48, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v48, vcc, v48, v23
; VI-NEXT:    v_add_u32_e32 v48, vcc, s6, v48
; VI-NEXT:    v_or_b32_e32 v49, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_cndmask_b32_e32 v23, v48, v49, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; VI-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
; VI-NEXT:    v_bfe_u32 v49, v48, 16, 1
; VI-NEXT:    v_add_u32_e32 v49, vcc, v49, v48
; VI-NEXT:    v_add_u32_e32 v49, vcc, s6, v49
; VI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; VI-NEXT:    v_or_b32_e32 v50, 0x400000, v48
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; VI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; VI-NEXT:    v_cndmask_b32_e32 v48, v49, v50, vcc
; VI-NEXT:    v_bfe_u32 v49, v24, 16, 1
; VI-NEXT:    v_add_u32_e32 v49, vcc, v49, v24
; VI-NEXT:    v_add_u32_e32 v49, vcc, s6, v49
; VI-NEXT:    v_or_b32_e32 v50, 0x400000, v24
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; VI-NEXT:    v_cndmask_b32_e32 v24, v49, v50, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v49, 16, v25
; VI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; VI-NEXT:    v_bfe_u32 v50, v49, 16, 1
; VI-NEXT:    v_add_u32_e32 v50, vcc, v50, v49
; VI-NEXT:    v_add_u32_e32 v50, vcc, s6, v50
; VI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; VI-NEXT:    v_or_b32_e32 v51, 0x400000, v49
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; VI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; VI-NEXT:    v_cndmask_b32_e32 v49, v50, v51, vcc
; VI-NEXT:    v_bfe_u32 v50, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v50, vcc, v50, v25
; VI-NEXT:    v_add_u32_e32 v50, vcc, s6, v50
; VI-NEXT:    v_or_b32_e32 v51, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_cndmask_b32_e32 v25, v50, v51, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v50, 16, v26
; VI-NEXT:    v_add_f32_e32 v50, 0x40c00000, v50
; VI-NEXT:    v_bfe_u32 v51, v50, 16, 1
; VI-NEXT:    v_add_u32_e32 v51, vcc, v51, v50
; VI-NEXT:    v_add_u32_e32 v51, vcc, s6, v51
; VI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; VI-NEXT:    v_or_b32_e32 v52, 0x400000, v50
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; VI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; VI-NEXT:    v_cndmask_b32_e32 v50, v51, v52, vcc
; VI-NEXT:    v_bfe_u32 v51, v26, 16, 1
; VI-NEXT:    v_add_u32_e32 v51, vcc, v51, v26
; VI-NEXT:    v_add_u32_e32 v51, vcc, s6, v51
; VI-NEXT:    v_or_b32_e32 v52, 0x400000, v26
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; VI-NEXT:    v_cndmask_b32_e32 v26, v51, v52, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v51, 16, v27
; VI-NEXT:    v_add_f32_e32 v51, 0x40c00000, v51
; VI-NEXT:    v_bfe_u32 v52, v51, 16, 1
; VI-NEXT:    v_add_u32_e32 v52, vcc, v52, v51
; VI-NEXT:    v_add_u32_e32 v52, vcc, s6, v52
; VI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; VI-NEXT:    v_or_b32_e32 v53, 0x400000, v51
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; VI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; VI-NEXT:    v_cndmask_b32_e32 v51, v52, v53, vcc
; VI-NEXT:    v_bfe_u32 v52, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v52, vcc, v52, v27
; VI-NEXT:    v_add_u32_e32 v52, vcc, s6, v52
; VI-NEXT:    v_or_b32_e32 v53, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_cndmask_b32_e32 v27, v52, v53, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v52, 16, v28
; VI-NEXT:    v_add_f32_e32 v52, 0x40c00000, v52
; VI-NEXT:    v_bfe_u32 v53, v52, 16, 1
; VI-NEXT:    v_add_u32_e32 v53, vcc, v53, v52
; VI-NEXT:    v_add_u32_e32 v53, vcc, s6, v53
; VI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; VI-NEXT:    v_or_b32_e32 v54, 0x400000, v52
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; VI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; VI-NEXT:    v_cndmask_b32_e32 v52, v53, v54, vcc
; VI-NEXT:    v_bfe_u32 v53, v28, 16, 1
; VI-NEXT:    v_add_u32_e32 v53, vcc, v53, v28
; VI-NEXT:    v_add_u32_e32 v53, vcc, s6, v53
; VI-NEXT:    v_or_b32_e32 v54, 0x400000, v28
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; VI-NEXT:    v_cndmask_b32_e32 v28, v53, v54, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v53, 16, v29
; VI-NEXT:    v_add_f32_e32 v53, 0x40c00000, v53
; VI-NEXT:    v_bfe_u32 v54, v53, 16, 1
; VI-NEXT:    v_add_u32_e32 v54, vcc, v54, v53
; VI-NEXT:    v_add_u32_e32 v54, vcc, s6, v54
; VI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; VI-NEXT:    v_or_b32_e32 v55, 0x400000, v53
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; VI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; VI-NEXT:    v_cndmask_b32_e32 v53, v54, v55, vcc
; VI-NEXT:    v_bfe_u32 v54, v29, 16, 1
; VI-NEXT:    v_add_u32_e32 v54, vcc, v54, v29
; VI-NEXT:    v_add_u32_e32 v54, vcc, s6, v54
; VI-NEXT:    v_or_b32_e32 v55, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_cndmask_b32_e32 v29, v54, v55, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v54, 16, v30
; VI-NEXT:    v_add_f32_e32 v54, 0x40c00000, v54
; VI-NEXT:    v_bfe_u32 v55, v54, 16, 1
; VI-NEXT:    v_add_u32_e32 v55, vcc, v55, v54
; VI-NEXT:    v_add_u32_e32 v55, vcc, s6, v55
; VI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; VI-NEXT:    v_or_b32_e32 v40, 0x400000, v54
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; VI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; VI-NEXT:    v_cndmask_b32_e32 v54, v55, v40, vcc
; VI-NEXT:    v_bfe_u32 v55, v30, 16, 1
; VI-NEXT:    v_add_u32_e32 v55, vcc, v55, v30
; VI-NEXT:    v_add_u32_e32 v55, vcc, s6, v55
; VI-NEXT:    v_or_b32_e32 v40, 0x400000, v30
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; VI-NEXT:    v_cndmask_b32_e32 v30, v55, v40, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v55, 16, v31
; VI-NEXT:    v_add_f32_e32 v55, 0x40c00000, v55
; VI-NEXT:    v_bfe_u32 v40, v55, 16, 1
; VI-NEXT:    v_add_u32_e32 v40, vcc, v40, v55
; VI-NEXT:    v_add_u32_e32 v40, vcc, s6, v40
; VI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; VI-NEXT:    v_or_b32_e32 v41, 0x400000, v55
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; VI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; VI-NEXT:    v_cndmask_b32_e32 v55, v40, v41, vcc
; VI-NEXT:    v_bfe_u32 v40, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v40, vcc, v40, v31
; VI-NEXT:    v_add_u32_e32 v40, vcc, s6, v40
; VI-NEXT:    v_or_b32_e32 v41, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v40, v41, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v40, 16, v0
; VI-NEXT:    v_add_f32_e32 v40, 0x40c00000, v40
; VI-NEXT:    v_bfe_u32 v41, v40, 16, 1
; VI-NEXT:    v_add_u32_e32 v41, vcc, v41, v40
; VI-NEXT:    v_add_u32_e32 v41, vcc, s6, v41
; VI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; VI-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; VI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; VI-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; VI-NEXT:    v_bfe_u32 v41, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v41, vcc, v41, v0
; VI-NEXT:    v_add_u32_e32 v41, vcc, s6, v41
; VI-NEXT:    v_or_b32_e32 v42, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v17, v41, v42, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v41, 16, v1
; VI-NEXT:    v_add_f32_e32 v41, 0x40c00000, v41
; VI-NEXT:    v_bfe_u32 v42, v41, 16, 1
; VI-NEXT:    v_add_u32_e32 v42, vcc, v42, v41
; VI-NEXT:    v_add_u32_e32 v42, vcc, s6, v42
; VI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; VI-NEXT:    v_or_b32_e32 v43, 0x400000, v41
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; VI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; VI-NEXT:    v_cndmask_b32_e32 v41, v42, v43, vcc
; VI-NEXT:    v_bfe_u32 v42, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v42, vcc, v42, v1
; VI-NEXT:    v_add_u32_e32 v42, vcc, s6, v42
; VI-NEXT:    v_or_b32_e32 v43, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_cndmask_b32_e32 v16, v42, v43, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v42, 16, v2
; VI-NEXT:    v_add_f32_e32 v42, 0x40c00000, v42
; VI-NEXT:    v_bfe_u32 v43, v42, 16, 1
; VI-NEXT:    v_add_u32_e32 v43, vcc, v43, v42
; VI-NEXT:    v_add_u32_e32 v43, vcc, s6, v43
; VI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; VI-NEXT:    v_or_b32_e32 v44, 0x400000, v42
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v42, v42
; VI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; VI-NEXT:    v_cndmask_b32_e32 v42, v43, v44, vcc
; VI-NEXT:    v_bfe_u32 v43, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v43, vcc, v43, v2
; VI-NEXT:    v_add_u32_e32 v43, vcc, s6, v43
; VI-NEXT:    v_or_b32_e32 v44, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v43, v44, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v43, 16, v3
; VI-NEXT:    v_add_f32_e32 v43, 0x40c00000, v43
; VI-NEXT:    v_bfe_u32 v44, v43, 16, 1
; VI-NEXT:    v_add_u32_e32 v44, vcc, v44, v43
; VI-NEXT:    v_add_u32_e32 v44, vcc, s6, v44
; VI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; VI-NEXT:    v_or_b32_e32 v45, 0x400000, v43
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v43, v43
; VI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; VI-NEXT:    v_cndmask_b32_e32 v43, v44, v45, vcc
; VI-NEXT:    v_bfe_u32 v44, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v44, vcc, v44, v3
; VI-NEXT:    v_add_u32_e32 v44, vcc, s6, v44
; VI-NEXT:    v_or_b32_e32 v45, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v44, v45, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; VI-NEXT:    v_add_f32_e32 v44, 0x40c00000, v44
; VI-NEXT:    v_bfe_u32 v45, v44, 16, 1
; VI-NEXT:    v_add_u32_e32 v45, vcc, v45, v44
; VI-NEXT:    v_add_u32_e32 v45, vcc, s6, v45
; VI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; VI-NEXT:    v_or_b32_e32 v46, 0x400000, v44
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v44, v44
; VI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; VI-NEXT:    v_cndmask_b32_e32 v44, v45, v46, vcc
; VI-NEXT:    v_bfe_u32 v45, v4, 16, 1
; VI-NEXT:    v_add_u32_e32 v45, vcc, v45, v4
; VI-NEXT:    v_add_u32_e32 v45, vcc, s6, v45
; VI-NEXT:    v_or_b32_e32 v46, 0x400000, v4
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; VI-NEXT:    v_cndmask_b32_e32 v4, v45, v46, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v45, 16, v5
; VI-NEXT:    v_add_f32_e32 v45, 0x40c00000, v45
; VI-NEXT:    v_bfe_u32 v46, v45, 16, 1
; VI-NEXT:    v_add_u32_e32 v46, vcc, v46, v45
; VI-NEXT:    v_add_u32_e32 v46, vcc, s6, v46
; VI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; VI-NEXT:    v_or_b32_e32 v47, 0x400000, v45
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v45, v45
; VI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; VI-NEXT:    v_cndmask_b32_e32 v45, v46, v47, vcc
; VI-NEXT:    v_bfe_u32 v46, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v46, vcc, v46, v5
; VI-NEXT:    v_add_u32_e32 v46, vcc, s6, v46
; VI-NEXT:    v_or_b32_e32 v47, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v46, v47, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v46, 16, v6
; VI-NEXT:    v_add_f32_e32 v46, 0x40c00000, v46
; VI-NEXT:    v_bfe_u32 v47, v46, 16, 1
; VI-NEXT:    v_add_u32_e32 v47, vcc, v47, v46
; VI-NEXT:    v_add_u32_e32 v47, vcc, s6, v47
; VI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; VI-NEXT:    v_or_b32_e32 v56, 0x400000, v46
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v46, v46
; VI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; VI-NEXT:    v_cndmask_b32_e32 v46, v47, v56, vcc
; VI-NEXT:    v_bfe_u32 v47, v6, 16, 1
; VI-NEXT:    v_add_u32_e32 v47, vcc, v47, v6
; VI-NEXT:    v_add_u32_e32 v47, vcc, s6, v47
; VI-NEXT:    v_or_b32_e32 v56, 0x400000, v6
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; VI-NEXT:    v_cndmask_b32_e32 v6, v47, v56, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v47, 16, v7
; VI-NEXT:    v_add_f32_e32 v47, 0x40c00000, v47
; VI-NEXT:    v_bfe_u32 v56, v47, 16, 1
; VI-NEXT:    v_add_u32_e32 v56, vcc, v56, v47
; VI-NEXT:    v_add_u32_e32 v56, vcc, s6, v56
; VI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; VI-NEXT:    v_or_b32_e32 v57, 0x400000, v47
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v47, v47
; VI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; VI-NEXT:    v_cndmask_b32_e32 v47, v56, v57, vcc
; VI-NEXT:    v_bfe_u32 v56, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v56, vcc, v56, v7
; VI-NEXT:    v_add_u32_e32 v56, vcc, s6, v56
; VI-NEXT:    v_or_b32_e32 v57, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v56, v57, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v56, 16, v8
; VI-NEXT:    v_add_f32_e32 v56, 0x40c00000, v56
; VI-NEXT:    v_bfe_u32 v57, v56, 16, 1
; VI-NEXT:    v_add_u32_e32 v57, vcc, v57, v56
; VI-NEXT:    v_add_u32_e32 v57, vcc, s6, v57
; VI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; VI-NEXT:    v_or_b32_e32 v58, 0x400000, v56
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v56, v56
; VI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; VI-NEXT:    v_cndmask_b32_e32 v56, v57, v58, vcc
; VI-NEXT:    v_bfe_u32 v57, v8, 16, 1
; VI-NEXT:    v_add_u32_e32 v57, vcc, v57, v8
; VI-NEXT:    v_add_u32_e32 v57, vcc, s6, v57
; VI-NEXT:    v_or_b32_e32 v58, 0x400000, v8
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; VI-NEXT:    v_cndmask_b32_e32 v8, v57, v58, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v57, 16, v9
; VI-NEXT:    v_add_f32_e32 v57, 0x40c00000, v57
; VI-NEXT:    v_bfe_u32 v58, v57, 16, 1
; VI-NEXT:    v_add_u32_e32 v58, vcc, v58, v57
; VI-NEXT:    v_add_u32_e32 v58, vcc, s6, v58
; VI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; VI-NEXT:    v_or_b32_e32 v59, 0x400000, v57
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v57, v57
; VI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; VI-NEXT:    v_cndmask_b32_e32 v57, v58, v59, vcc
; VI-NEXT:    v_bfe_u32 v58, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v58, vcc, v58, v9
; VI-NEXT:    v_add_u32_e32 v58, vcc, s6, v58
; VI-NEXT:    v_or_b32_e32 v59, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v58, v59, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v58, 16, v10
; VI-NEXT:    v_add_f32_e32 v58, 0x40c00000, v58
; VI-NEXT:    v_bfe_u32 v59, v58, 16, 1
; VI-NEXT:    v_add_u32_e32 v59, vcc, v59, v58
; VI-NEXT:    v_add_u32_e32 v59, vcc, s6, v59
; VI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; VI-NEXT:    v_or_b32_e32 v60, 0x400000, v58
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v58, v58
; VI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; VI-NEXT:    v_cndmask_b32_e32 v58, v59, v60, vcc
; VI-NEXT:    v_bfe_u32 v59, v10, 16, 1
; VI-NEXT:    v_add_u32_e32 v59, vcc, v59, v10
; VI-NEXT:    v_add_u32_e32 v59, vcc, s6, v59
; VI-NEXT:    v_or_b32_e32 v60, 0x400000, v10
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; VI-NEXT:    v_cndmask_b32_e32 v10, v59, v60, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
; VI-NEXT:    v_add_f32_e32 v59, 0x40c00000, v59
; VI-NEXT:    v_bfe_u32 v60, v59, 16, 1
; VI-NEXT:    v_add_u32_e32 v60, vcc, v60, v59
; VI-NEXT:    v_add_u32_e32 v60, vcc, s6, v60
; VI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; VI-NEXT:    v_or_b32_e32 v61, 0x400000, v59
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v59, v59
; VI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; VI-NEXT:    v_cndmask_b32_e32 v59, v60, v61, vcc
; VI-NEXT:    v_bfe_u32 v60, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v60, vcc, v60, v11
; VI-NEXT:    v_add_u32_e32 v60, vcc, s6, v60
; VI-NEXT:    v_or_b32_e32 v61, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v60, v61, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v60, 16, v12
; VI-NEXT:    v_add_f32_e32 v60, 0x40c00000, v60
; VI-NEXT:    v_bfe_u32 v61, v60, 16, 1
; VI-NEXT:    v_add_u32_e32 v61, vcc, v61, v60
; VI-NEXT:    v_add_u32_e32 v61, vcc, s6, v61
; VI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; VI-NEXT:    v_or_b32_e32 v62, 0x400000, v60
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v60, v60
; VI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; VI-NEXT:    v_cndmask_b32_e32 v60, v61, v62, vcc
; VI-NEXT:    v_bfe_u32 v61, v12, 16, 1
; VI-NEXT:    v_add_u32_e32 v61, vcc, v61, v12
; VI-NEXT:    v_add_u32_e32 v61, vcc, s6, v61
; VI-NEXT:    v_or_b32_e32 v62, 0x400000, v12
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; VI-NEXT:    v_cndmask_b32_e32 v12, v61, v62, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v61, 16, v13
; VI-NEXT:    v_add_f32_e32 v61, 0x40c00000, v61
; VI-NEXT:    v_bfe_u32 v62, v61, 16, 1
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v61
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v61
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v61, v61
; VI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; VI-NEXT:    v_cndmask_b32_e32 v61, v62, v63, vcc
; VI-NEXT:    v_bfe_u32 v62, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v13
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v62, v63, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v62, 16, v14
; VI-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; VI-NEXT:    v_bfe_u32 v63, v62, 16, 1
; VI-NEXT:    v_add_u32_e32 v63, vcc, v63, v62
; VI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; VI-NEXT:    v_add_u32_e32 v63, vcc, s6, v63
; VI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; VI-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; VI-NEXT:    v_bfe_u32 v62, v14, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v1, v63, v0, vcc
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v14
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v14
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; VI-NEXT:    v_cndmask_b32_e32 v14, v62, v63, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v62, 16, v15
; VI-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; VI-NEXT:    v_bfe_u32 v63, v62, 16, 1
; VI-NEXT:    v_add_u32_e32 v63, vcc, v63, v62
; VI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; VI-NEXT:    v_add_u32_e32 v63, vcc, s6, v63
; VI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; VI-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; VI-NEXT:    v_bfe_u32 v62, v15, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v0, v63, v0, vcc
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v15
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v62, v63, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_alignbit_b32 v15, v15, v0, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v14
; VI-NEXT:    v_alignbit_b32 v14, v0, v1, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v13
; VI-NEXT:    v_alignbit_b32 v13, v0, v61, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v12
; VI-NEXT:    v_alignbit_b32 v12, v0, v60, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v11
; VI-NEXT:    v_alignbit_b32 v11, v0, v59, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v10
; VI-NEXT:    v_alignbit_b32 v10, v0, v58, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v9
; VI-NEXT:    v_alignbit_b32 v9, v0, v57, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v8
; VI-NEXT:    v_alignbit_b32 v8, v0, v56, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v7
; VI-NEXT:    v_alignbit_b32 v7, v0, v47, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v6
; VI-NEXT:    v_alignbit_b32 v6, v0, v46, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v5
; VI-NEXT:    v_alignbit_b32 v5, v0, v45, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v4
; VI-NEXT:    v_alignbit_b32 v4, v0, v44, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v3
; VI-NEXT:    v_alignbit_b32 v3, v0, v43, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; VI-NEXT:    v_alignbit_b32 v2, v0, v42, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v31
; VI-NEXT:    v_alignbit_b32 v31, v16, v55, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v30
; VI-NEXT:    v_alignbit_b32 v30, v16, v54, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v29
; VI-NEXT:    v_alignbit_b32 v29, v16, v53, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v28
; VI-NEXT:    v_alignbit_b32 v28, v16, v52, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v27
; VI-NEXT:    v_alignbit_b32 v27, v16, v51, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v26
; VI-NEXT:    v_alignbit_b32 v26, v16, v50, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v25
; VI-NEXT:    v_alignbit_b32 v25, v16, v49, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v24
; VI-NEXT:    v_alignbit_b32 v24, v16, v48, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v23
; VI-NEXT:    v_alignbit_b32 v23, v16, v39, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v22
; VI-NEXT:    v_alignbit_b32 v22, v16, v38, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v21
; VI-NEXT:    v_alignbit_b32 v21, v16, v37, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
; VI-NEXT:    v_alignbit_b32 v20, v16, v36, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v19
; VI-NEXT:    v_alignbit_b32 v19, v16, v35, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; VI-NEXT:    v_alignbit_b32 v18, v16, v34, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v32
; VI-NEXT:    v_alignbit_b32 v1, v0, v41, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v17
; VI-NEXT:    v_alignbit_b32 v17, v16, v33, 16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    v_alignbit_b32 v0, v0, v40, 16
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_alignbit_b32 v16, v16, v32, 16
; VI-NEXT:  .LBB100_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v64f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB100_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    s_movk_i32 s6, 0x7fff
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v16, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT:    v_add3_u32 v34, v34, v33, s6
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v34, v34, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v34, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX9-NEXT:    v_bfe_u32 v35, v34, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT:    v_add3_u32 v35, v35, v34, s6
; GFX9-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc
; GFX9-NEXT:    v_bfe_u32 v35, v18, 16, 1
; GFX9-NEXT:    v_add3_u32 v35, v35, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v35, v36, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
; GFX9-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX9-NEXT:    v_bfe_u32 v36, v35, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT:    v_add3_u32 v36, v36, v35, s6
; GFX9-NEXT:    v_or_b32_e32 v37, 0x400000, v35
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX9-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v35, v36, v37, vcc
; GFX9-NEXT:    v_bfe_u32 v36, v19, 16, 1
; GFX9-NEXT:    v_add3_u32 v36, v36, v19, s6
; GFX9-NEXT:    v_or_b32_e32 v37, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v36, v37, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
; GFX9-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
; GFX9-NEXT:    v_bfe_u32 v37, v36, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT:    v_add3_u32 v37, v37, v36, s6
; GFX9-NEXT:    v_or_b32_e32 v38, 0x400000, v36
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX9-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v36, v37, v38, vcc
; GFX9-NEXT:    v_bfe_u32 v37, v20, 16, 1
; GFX9-NEXT:    v_add3_u32 v37, v37, v20, s6
; GFX9-NEXT:    v_or_b32_e32 v38, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v37, v38, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; GFX9-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
; GFX9-NEXT:    v_bfe_u32 v38, v37, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT:    v_add3_u32 v38, v38, v37, s6
; GFX9-NEXT:    v_or_b32_e32 v39, 0x400000, v37
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX9-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v37, v38, v39, vcc
; GFX9-NEXT:    v_bfe_u32 v38, v21, 16, 1
; GFX9-NEXT:    v_add3_u32 v38, v38, v21, s6
; GFX9-NEXT:    v_or_b32_e32 v39, 0x400000, v21
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v21, v38, v39, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v38, 16, v22
; GFX9-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; GFX9-NEXT:    v_bfe_u32 v39, v38, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT:    v_add3_u32 v39, v39, v38, s6
; GFX9-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX9-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v38, v39, v48, vcc
; GFX9-NEXT:    v_bfe_u32 v39, v22, 16, 1
; GFX9-NEXT:    v_add3_u32 v39, v39, v22, s6
; GFX9-NEXT:    v_or_b32_e32 v48, 0x400000, v22
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v22, v39, v48, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
; GFX9-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
; GFX9-NEXT:    v_bfe_u32 v48, v39, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT:    v_add3_u32 v48, v48, v39, s6
; GFX9-NEXT:    v_or_b32_e32 v49, 0x400000, v39
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX9-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v39, v48, v49, vcc
; GFX9-NEXT:    v_bfe_u32 v48, v23, 16, 1
; GFX9-NEXT:    v_add3_u32 v48, v48, v23, s6
; GFX9-NEXT:    v_or_b32_e32 v49, 0x400000, v23
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v23, v48, v49, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; GFX9-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
; GFX9-NEXT:    v_bfe_u32 v49, v48, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT:    v_add3_u32 v49, v49, v48, s6
; GFX9-NEXT:    v_or_b32_e32 v50, 0x400000, v48
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX9-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v48, v49, v50, vcc
; GFX9-NEXT:    v_bfe_u32 v49, v24, 16, 1
; GFX9-NEXT:    v_add3_u32 v49, v49, v24, s6
; GFX9-NEXT:    v_or_b32_e32 v50, 0x400000, v24
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v24, v49, v50, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v49, 16, v25
; GFX9-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; GFX9-NEXT:    v_bfe_u32 v50, v49, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT:    v_add3_u32 v50, v50, v49, s6
; GFX9-NEXT:    v_or_b32_e32 v51, 0x400000, v49
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX9-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v49, v50, v51, vcc
; GFX9-NEXT:    v_bfe_u32 v50, v25, 16, 1
; GFX9-NEXT:    v_add3_u32 v50, v50, v25, s6
; GFX9-NEXT:    v_or_b32_e32 v51, 0x400000, v25
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v25, v50, v51, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v50, 16, v26
; GFX9-NEXT:    v_add_f32_e32 v50, 0x40c00000, v50
; GFX9-NEXT:    v_bfe_u32 v51, v50, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT:    v_add3_u32 v51, v51, v50, s6
; GFX9-NEXT:    v_or_b32_e32 v52, 0x400000, v50
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX9-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v50, v51, v52, vcc
; GFX9-NEXT:    v_bfe_u32 v51, v26, 16, 1
; GFX9-NEXT:    v_add3_u32 v51, v51, v26, s6
; GFX9-NEXT:    v_or_b32_e32 v52, 0x400000, v26
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v26, v51, v52, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v51, 16, v27
; GFX9-NEXT:    v_add_f32_e32 v51, 0x40c00000, v51
; GFX9-NEXT:    v_bfe_u32 v52, v51, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT:    v_add3_u32 v52, v52, v51, s6
; GFX9-NEXT:    v_or_b32_e32 v53, 0x400000, v51
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX9-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v51, v52, v53, vcc
; GFX9-NEXT:    v_bfe_u32 v52, v27, 16, 1
; GFX9-NEXT:    v_add3_u32 v52, v52, v27, s6
; GFX9-NEXT:    v_or_b32_e32 v53, 0x400000, v27
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v27, v52, v53, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v52, 16, v28
; GFX9-NEXT:    v_add_f32_e32 v52, 0x40c00000, v52
; GFX9-NEXT:    v_bfe_u32 v53, v52, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT:    v_add3_u32 v53, v53, v52, s6
; GFX9-NEXT:    v_or_b32_e32 v54, 0x400000, v52
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX9-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v52, v53, v54, vcc
; GFX9-NEXT:    v_bfe_u32 v53, v28, 16, 1
; GFX9-NEXT:    v_add3_u32 v53, v53, v28, s6
; GFX9-NEXT:    v_or_b32_e32 v54, 0x400000, v28
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v28, v53, v54, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v53, 16, v29
; GFX9-NEXT:    v_add_f32_e32 v53, 0x40c00000, v53
; GFX9-NEXT:    v_bfe_u32 v54, v53, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT:    v_add3_u32 v54, v54, v53, s6
; GFX9-NEXT:    v_or_b32_e32 v55, 0x400000, v53
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX9-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v53, v54, v55, vcc
; GFX9-NEXT:    v_bfe_u32 v54, v29, 16, 1
; GFX9-NEXT:    v_add3_u32 v54, v54, v29, s6
; GFX9-NEXT:    v_or_b32_e32 v55, 0x400000, v29
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v29, v54, v55, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v54, 16, v30
; GFX9-NEXT:    v_add_f32_e32 v54, 0x40c00000, v54
; GFX9-NEXT:    v_bfe_u32 v55, v54, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT:    v_add3_u32 v55, v55, v54, s6
; GFX9-NEXT:    v_or_b32_e32 v40, 0x400000, v54
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX9-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v54, v55, v40, vcc
; GFX9-NEXT:    v_bfe_u32 v55, v30, 16, 1
; GFX9-NEXT:    v_add3_u32 v55, v55, v30, s6
; GFX9-NEXT:    v_or_b32_e32 v40, 0x400000, v30
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v30, v55, v40, vcc
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_lshlrev_b32_e32 v55, 16, v31
; GFX9-NEXT:    v_add_f32_e32 v55, 0x40c00000, v55
; GFX9-NEXT:    v_bfe_u32 v40, v55, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX9-NEXT:    v_add3_u32 v40, v40, v55, s6
; GFX9-NEXT:    v_or_b32_e32 v41, 0x400000, v55
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v55, v40, v41, vcc
; GFX9-NEXT:    v_bfe_u32 v40, v31, 16, 1
; GFX9-NEXT:    v_add3_u32 v40, v40, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v41, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v31, v40, v41, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v40, 16, v0
; GFX9-NEXT:    v_add_f32_e32 v40, 0x40c00000, v40
; GFX9-NEXT:    v_bfe_u32 v41, v40, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT:    v_add3_u32 v41, v41, v40, s6
; GFX9-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX9-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; GFX9-NEXT:    v_bfe_u32 v41, v0, 16, 1
; GFX9-NEXT:    v_add3_u32 v41, v41, v0, s6
; GFX9-NEXT:    v_or_b32_e32 v42, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v41, v42, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v41, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v41, 0x40c00000, v41
; GFX9-NEXT:    v_bfe_u32 v42, v41, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT:    v_add3_u32 v42, v42, v41, s6
; GFX9-NEXT:    v_or_b32_e32 v43, 0x400000, v41
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v41, v42, v43, vcc
; GFX9-NEXT:    v_bfe_u32 v42, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v42, v42, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v43, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v42, v43, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v42, 16, v2
; GFX9-NEXT:    v_add_f32_e32 v42, 0x40c00000, v42
; GFX9-NEXT:    v_bfe_u32 v43, v42, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT:    v_add3_u32 v43, v43, v42, s6
; GFX9-NEXT:    v_or_b32_e32 v44, 0x400000, v42
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v42, v42
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v42, v43, v44, vcc
; GFX9-NEXT:    v_bfe_u32 v43, v2, 16, 1
; GFX9-NEXT:    v_add3_u32 v43, v43, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v44, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v43, v44, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v43, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v43, 0x40c00000, v43
; GFX9-NEXT:    v_bfe_u32 v44, v43, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT:    v_add3_u32 v44, v44, v43, s6
; GFX9-NEXT:    v_or_b32_e32 v45, 0x400000, v43
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v43, v43
; GFX9-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v43, v44, v45, vcc
; GFX9-NEXT:    v_bfe_u32 v44, v3, 16, 1
; GFX9-NEXT:    v_add3_u32 v44, v44, v3, s6
; GFX9-NEXT:    v_or_b32_e32 v45, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v44, v45, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; GFX9-NEXT:    v_add_f32_e32 v44, 0x40c00000, v44
; GFX9-NEXT:    v_bfe_u32 v45, v44, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT:    v_add3_u32 v45, v45, v44, s6
; GFX9-NEXT:    v_or_b32_e32 v46, 0x400000, v44
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v44, v44
; GFX9-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v44, v45, v46, vcc
; GFX9-NEXT:    v_bfe_u32 v45, v4, 16, 1
; GFX9-NEXT:    v_add3_u32 v45, v45, v4, s6
; GFX9-NEXT:    v_or_b32_e32 v46, 0x400000, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v4, v45, v46, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v45, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v45, 0x40c00000, v45
; GFX9-NEXT:    v_bfe_u32 v46, v45, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT:    v_add3_u32 v46, v46, v45, s6
; GFX9-NEXT:    v_or_b32_e32 v47, 0x400000, v45
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v45, v45
; GFX9-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v45, v46, v47, vcc
; GFX9-NEXT:    v_bfe_u32 v46, v5, 16, 1
; GFX9-NEXT:    v_add3_u32 v46, v46, v5, s6
; GFX9-NEXT:    v_or_b32_e32 v47, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v46, v47, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v46, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v46, 0x40c00000, v46
; GFX9-NEXT:    v_bfe_u32 v47, v46, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT:    v_add3_u32 v47, v47, v46, s6
; GFX9-NEXT:    v_or_b32_e32 v56, 0x400000, v46
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v46, v46
; GFX9-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v46, v47, v56, vcc
; GFX9-NEXT:    v_bfe_u32 v47, v6, 16, 1
; GFX9-NEXT:    v_add3_u32 v47, v47, v6, s6
; GFX9-NEXT:    v_or_b32_e32 v56, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v47, v56, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v47, 16, v7
; GFX9-NEXT:    v_add_f32_e32 v47, 0x40c00000, v47
; GFX9-NEXT:    v_bfe_u32 v56, v47, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT:    v_add3_u32 v56, v56, v47, s6
; GFX9-NEXT:    v_or_b32_e32 v57, 0x400000, v47
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v47, v47
; GFX9-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v47, v56, v57, vcc
; GFX9-NEXT:    v_bfe_u32 v56, v7, 16, 1
; GFX9-NEXT:    v_add3_u32 v56, v56, v7, s6
; GFX9-NEXT:    v_or_b32_e32 v57, 0x400000, v7
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v7, v56, v57, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v56, 16, v8
; GFX9-NEXT:    v_add_f32_e32 v56, 0x40c00000, v56
; GFX9-NEXT:    v_bfe_u32 v57, v56, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT:    v_add3_u32 v57, v57, v56, s6
; GFX9-NEXT:    v_or_b32_e32 v58, 0x400000, v56
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v56, v56
; GFX9-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v56, v57, v58, vcc
; GFX9-NEXT:    v_bfe_u32 v57, v8, 16, 1
; GFX9-NEXT:    v_add3_u32 v57, v57, v8, s6
; GFX9-NEXT:    v_or_b32_e32 v58, 0x400000, v8
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v8, v57, v58, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v57, 16, v9
; GFX9-NEXT:    v_add_f32_e32 v57, 0x40c00000, v57
; GFX9-NEXT:    v_bfe_u32 v58, v57, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT:    v_add3_u32 v58, v58, v57, s6
; GFX9-NEXT:    v_or_b32_e32 v59, 0x400000, v57
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v57, v57
; GFX9-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v57, v58, v59, vcc
; GFX9-NEXT:    v_bfe_u32 v58, v9, 16, 1
; GFX9-NEXT:    v_add3_u32 v58, v58, v9, s6
; GFX9-NEXT:    v_or_b32_e32 v59, 0x400000, v9
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v9, v58, v59, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v58, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v58, 0x40c00000, v58
; GFX9-NEXT:    v_bfe_u32 v59, v58, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT:    v_add3_u32 v59, v59, v58, s6
; GFX9-NEXT:    v_or_b32_e32 v60, 0x400000, v58
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v58, v58
; GFX9-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v58, v59, v60, vcc
; GFX9-NEXT:    v_bfe_u32 v59, v10, 16, 1
; GFX9-NEXT:    v_add3_u32 v59, v59, v10, s6
; GFX9-NEXT:    v_or_b32_e32 v60, 0x400000, v10
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v10, v59, v60, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
; GFX9-NEXT:    v_add_f32_e32 v59, 0x40c00000, v59
; GFX9-NEXT:    v_bfe_u32 v60, v59, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT:    v_add3_u32 v60, v60, v59, s6
; GFX9-NEXT:    v_or_b32_e32 v61, 0x400000, v59
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v59, v59
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v59, v60, v61, vcc
; GFX9-NEXT:    v_bfe_u32 v60, v11, 16, 1
; GFX9-NEXT:    v_add3_u32 v60, v60, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v61, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v60, v61, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v60, 16, v12
; GFX9-NEXT:    v_add_f32_e32 v60, 0x40c00000, v60
; GFX9-NEXT:    v_bfe_u32 v61, v60, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT:    v_add3_u32 v61, v61, v60, s6
; GFX9-NEXT:    v_or_b32_e32 v62, 0x400000, v60
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v60, v60
; GFX9-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v60, v61, v62, vcc
; GFX9-NEXT:    v_bfe_u32 v61, v12, 16, 1
; GFX9-NEXT:    v_add3_u32 v61, v61, v12, s6
; GFX9-NEXT:    v_or_b32_e32 v62, 0x400000, v12
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v12, v61, v62, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v61, 16, v13
; GFX9-NEXT:    v_add_f32_e32 v61, 0x40c00000, v61
; GFX9-NEXT:    v_bfe_u32 v62, v61, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT:    v_add3_u32 v62, v62, v61, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v61
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v61, v61
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v61, v62, v63, vcc
; GFX9-NEXT:    v_bfe_u32 v62, v13, 16, 1
; GFX9-NEXT:    v_add3_u32 v62, v62, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v62, v63, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v62, 16, v14
; GFX9-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT:    v_bfe_u32 v63, v62, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX9-NEXT:    v_add3_u32 v63, v63, v62, s6
; GFX9-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; GFX9-NEXT:    v_bfe_u32 v62, v14, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v63, v0, vcc
; GFX9-NEXT:    v_add3_u32 v62, v62, v14, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v14
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v14, v62, v63, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v62, 16, v15
; GFX9-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT:    v_bfe_u32 v63, v62, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX9-NEXT:    v_add3_u32 v63, v63, v62, s6
; GFX9-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; GFX9-NEXT:    v_bfe_u32 v62, v15, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v63, v0, vcc
; GFX9-NEXT:    v_add3_u32 v62, v62, v15, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v15
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v62, v63, vcc
; GFX9-NEXT:    s_mov_b32 s6, 0x7060302
; GFX9-NEXT:    v_perm_b32 v15, v15, v0, s6
; GFX9-NEXT:    v_perm_b32 v14, v14, v1, s6
; GFX9-NEXT:    v_perm_b32 v1, v16, v41, s6
; GFX9-NEXT:    v_perm_b32 v0, v17, v40, s6
; GFX9-NEXT:    v_perm_b32 v17, v32, v33, s6
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v13, v13, v61, s6
; GFX9-NEXT:    v_perm_b32 v12, v12, v60, s6
; GFX9-NEXT:    v_perm_b32 v11, v11, v59, s6
; GFX9-NEXT:    v_perm_b32 v10, v10, v58, s6
; GFX9-NEXT:    v_perm_b32 v9, v9, v57, s6
; GFX9-NEXT:    v_perm_b32 v8, v8, v56, s6
; GFX9-NEXT:    v_perm_b32 v7, v7, v47, s6
; GFX9-NEXT:    v_perm_b32 v6, v6, v46, s6
; GFX9-NEXT:    v_perm_b32 v5, v5, v45, s6
; GFX9-NEXT:    v_perm_b32 v4, v4, v44, s6
; GFX9-NEXT:    v_perm_b32 v3, v3, v43, s6
; GFX9-NEXT:    v_perm_b32 v2, v2, v42, s6
; GFX9-NEXT:    v_perm_b32 v31, v31, v55, s6
; GFX9-NEXT:    v_perm_b32 v30, v30, v54, s6
; GFX9-NEXT:    v_perm_b32 v29, v29, v53, s6
; GFX9-NEXT:    v_perm_b32 v28, v28, v52, s6
; GFX9-NEXT:    v_perm_b32 v27, v27, v51, s6
; GFX9-NEXT:    v_perm_b32 v26, v26, v50, s6
; GFX9-NEXT:    v_perm_b32 v25, v25, v49, s6
; GFX9-NEXT:    v_perm_b32 v24, v24, v48, s6
; GFX9-NEXT:    v_perm_b32 v23, v23, v39, s6
; GFX9-NEXT:    v_perm_b32 v22, v22, v38, s6
; GFX9-NEXT:    v_perm_b32 v21, v21, v37, s6
; GFX9-NEXT:    v_perm_b32 v20, v20, v36, s6
; GFX9-NEXT:    v_perm_b32 v19, v19, v35, s6
; GFX9-NEXT:    v_perm_b32 v18, v18, v34, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v32, v16, s6
; GFX9-NEXT:  .LBB100_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v64f16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Spill
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 offset:68
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v41, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v42, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v43, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v44, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v45, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v46, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v47, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v56, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v57, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v58, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v59, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v60, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v61, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v62, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v63, s32 offset:8
; GFX11-TRUE16-NEXT:    v_mov_b32_e32 v48, v16
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v16, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v62, v30 :: v_dual_mov_b32 v61, v29
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v60, v28 :: v_dual_mov_b32 v59, v27
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v58, v26 :: v_dual_mov_b32 v57, v25
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v56, v24 :: v_dual_mov_b32 v55, v23
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v54, v22 :: v_dual_mov_b32 v53, v21
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v52, v20 :: v_dual_mov_b32 v51, v19
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v50, v18 :: v_dual_mov_b32 v49, v17
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr32_vgpr33_vgpr34_vgpr35_vgpr36_vgpr37_vgpr38_vgpr39_vgpr40_vgpr41_vgpr42_vgpr43_vgpr44_vgpr45_vgpr46_vgpr47
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v16
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31
; GFX11-TRUE16-NEXT:    s_and_saveexec_b32 s0, vcc_lo
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB100_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.false
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v47, v15 :: v_dual_mov_b32 v46, v14
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, v48 :: v_dual_mov_b32 v17, v49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v45, v13 :: v_dual_mov_b32 v44, v12
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v43, v11 :: v_dual_mov_b32 v42, v10
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v41, v9 :: v_dual_mov_b32 v40, v8
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v39, v7 :: v_dual_mov_b32 v38, v6
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v37, v5 :: v_dual_mov_b32 v36, v4
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v35, v3 :: v_dual_mov_b32 v34, v2
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v33, v1 :: v_dual_mov_b32 v32, v0
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, v50 :: v_dual_mov_b32 v19, v51
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, v52 :: v_dual_mov_b32 v21, v53
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, v54 :: v_dual_mov_b32 v23, v55
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, v56 :: v_dual_mov_b32 v25, v57
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, v58 :: v_dual_mov_b32 v27, v59
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, v60 :: v_dual_mov_b32 v29, v61
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, v62 :: v_dual_mov_b32 v31, v63
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr0
; GFX11-TRUE16-NEXT:    ; implicit-def: $vgpr48
; GFX11-TRUE16-NEXT:  .LBB100_2: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB100_4
; GFX11-TRUE16-NEXT:  ; %bb.3: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v49
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v17, 16, v48
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v50
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v54
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v56
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v32, 0xffff0000, v58
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v60
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff0000, v62
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v38, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v80, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v24, v24, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v48
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v49
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v84, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v27, 16, v53
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v22, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v29, 16, v55
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v21, v21, v16, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v57
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v59
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v16, v21, v22 :: v_dual_lshlrev_b32 v35, 16, v61
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v51
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v37, 16, v63
; GFX11-TRUE16-NEXT:    v_add3_u32 v23, v23, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v71, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v23, v25, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v21, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v18, 0x40c00000, v20 :: v_dual_lshlrev_b32 v25, 16, v51
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v50
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v21, v21, v19, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v24, v26, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v52
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v48.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v24, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v21, v23, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_bfe_u32 v23, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v49.h
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v19, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v19, v23, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v21, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v23, v24, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v25
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v20, v24, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v50, v19, v21, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v21, 0x40c00000, v26 :: v_dual_lshlrev_b32 v22, 16, v52
; GFX11-TRUE16-NEXT:    v_add3_u32 v20, v20, v24, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v23, v25, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v25, 0x400000, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v51, v20, v25, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v20, v26, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v53
; GFX11-TRUE16-NEXT:    v_bfe_u32 v25, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v51.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v20, v20, v24 :: v_dual_add_f32 v23, 0x40c00000, v23
; GFX11-TRUE16-NEXT:    v_add3_u32 v21, v25, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v24, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v52, v21, v24, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add3_u32 v25, v26, v23, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v23, 0x40c00000, v28 :: v_dual_lshlrev_b32 v24, 16, v54
; GFX11-TRUE16-NEXT:    v_bfe_u32 v22, v26, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v21, v25, v27, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v55
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v27, 0x400000, v26
; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v22, v22, v26, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v24, 0x40c00000, v24 :: v_dual_add_f32 v25, 0x40c00000, v25
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v50.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v53, v22, v27, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v22, v28, v23, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v27, v24, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v53.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v23, v27, v24, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v27, v28, v25, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v22, v26, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v26, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v29
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v24, v28, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v23, v26, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v25, 0x40c00000, v30 :: v_dual_lshlrev_b32 v26, 16, v56
; GFX11-TRUE16-NEXT:    v_add3_u32 v24, v24, v28, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v52.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v23, v27, v29, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v57
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v29, 0x400000, v28
; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v25
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v55, v24, v29, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v24, v30, v25, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v29, v26, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v27, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v55.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v25, v29, v26, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v24, v28, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v28, 0x400000, v26
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-TRUE16-NEXT:    v_add3_u32 v29, v30, v27, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v30, 0x40c00000, v31
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v64, v25, v28, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v26, v30, 16, 1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v27, 0x40c00000, v32 :: v_dual_lshlrev_b32 v28, 16, v58
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v54.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v25, v29, v31, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v59
; GFX11-TRUE16-NEXT:    v_add3_u32 v26, v26, v30, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v31, 0x400000, v30
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v27, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v28, 0x40c00000, v28 :: v_dual_add_f32 v29, 0x40c00000, v29
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, 0x400000, v27
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v65, v26, v31, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v26, v32, v27, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v31, v28, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v29, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v65.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v27, v31, v28, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v31, v32, v29, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v26, v30, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v30, 0x400000, v28
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v32, 0x40c00000, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v28, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v66, v27, v30, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v29, 0x40c00000, v34 :: v_dual_lshlrev_b32 v30, 16, v60
; GFX11-TRUE16-NEXT:    v_add3_u32 v28, v28, v32, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v64.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v27, v31, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v61
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v33, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v29, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v29
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v67, v28, v33, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v28, v34, v29, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v33, v30, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-TRUE16-NEXT:    v_bfe_u32 v34, v31, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v67.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v29, v33, v30, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v28, v28, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v32, 0x400000, v30
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-TRUE16-NEXT:    v_add3_u32 v33, v34, v31, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v35
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v31
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v68, v29, v32, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v30, v34, 16, 1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v31, 0x40c00000, v36 :: v_dual_lshlrev_b32 v32, 16, v62
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v66.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v29, v33, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v33, 0xffff0000, v63
; GFX11-TRUE16-NEXT:    v_add3_u32 v30, v30, v34, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v31, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_add_f32 v33, 0x40c00000, v33
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v69, v30, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v30, v36, v31, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v69.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v31, v35, v32, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v36, v33, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v30, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v37
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v33
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v32, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v70, v31, v34, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v33, 0x40c00000, v38
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v34, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v32, v36, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v0 :: v_dual_cndmask_b32 v31, v35, v37
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_bfe_u32 v36, v38, 16, 1
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v68.h
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v32, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v32, v37, v33, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v34, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_add3_u32 v33, v36, v38, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v0.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add3_u32 v36, v37, v34, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v32, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v35, 0x400000, v38
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v38, 0x400000, v37
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v33, v35, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v37, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v71
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v32.l, v1.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v37, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v36, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v36, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v34, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v34
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v36, 0x40c00000, v36 :: v_dual_lshlrev_b32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v70.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v71, v35, v38, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v39, v34, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v33.l, v71.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v35, v37, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v38, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v38, v39, v36, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v35, v37, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v80
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v37, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v35, v38, v39 :: v_dual_and_b32 v38, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v38, 0x40c00000, v38 :: v_dual_lshlrev_b32 v5, 16, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v37, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v80, v36, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_bfe_u32 v81, v38, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v35.l, v3.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v37, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v80, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v81, v38, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v38
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v37, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v82
; GFX11-TRUE16-NEXT:    v_add3_u32 v39, v39, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v37, v80, v81 :: v_dual_and_b32 v80, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v5
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v82, v38, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v80, 0x40c00000, v80 :: v_dual_lshlrev_b32 v7, 16, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v39, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v39, v82, v38, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v38
; GFX11-TRUE16-NEXT:    v_bfe_u32 v82, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-TRUE16-NEXT:    v_bfe_u32 v83, v80, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v37.l, v5.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v39, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v39, v82, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v82, v83, v80, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v80
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v39, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-TRUE16-NEXT:    v_bfe_u32 v81, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v80, 0x40c00000, v84 :: v_dual_cndmask_b32 v39, v82, v83
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v81, v81, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v80, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v82, 0x40c00000, v82 :: v_dual_lshlrev_b32 v9, 16, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v81, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v81, v84, v80, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v80
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v85, v82, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v39.l, v7.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v84, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v84, 0x400000, v82
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v40, v81, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v83, v85, v82, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v34.l, v2.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v80, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v81, 0x40c00000, v85
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v80, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v41, v83, v84, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v81, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v82, 0x40c00000, v82
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v40.l, v8.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v85, v82, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v80, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v84, v81, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v41.l, v9.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v42, v80, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v84, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v83, v85, v82, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v84, 0x400000, v82
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v80, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v81, 0x40c00000, v85
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v82, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v80, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v43, v83, v84, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v81, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v82, 0x40c00000, v82 :: v_dual_lshlrev_b32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v42.l, v10.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v36.l, v4.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v80, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v84, v81, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v85, v82, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v43.l, v11.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v44, v80, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v84, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v83, v85, v82, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v84, 0x400000, v82
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v85, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v80, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v81, 0x40c00000, v85 :: v_dual_lshlrev_b32 v82, 16, v15
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v80, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v45, v83, v84, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v81, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v82, 0x40c00000, v82
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v80, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_add3_u32 v83, v84, v81, 0x7fff
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v82, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v80, v14, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, 0x400000, v82
; GFX11-TRUE16-NEXT:    v_add3_u32 v84, v84, v82, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v87, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v80, v86, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v45.l, v13.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v87, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v44.l, v12.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v82, v84, v96, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v38.l, v6.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v46, v83, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v46.l, v14.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v47, v80, v86, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v47.l, v82.h
; GFX11-TRUE16-NEXT:  .LBB100_4: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, v40 :: v_dual_mov_b32 v9, v41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, v42 :: v_dual_mov_b32 v11, v43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, v44 :: v_dual_mov_b32 v13, v45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, v46 :: v_dual_mov_b32 v15, v47
; GFX11-TRUE16-NEXT:    s_clause 0xf ; 64-byte Folded Reload
; GFX11-TRUE16-NEXT:    scratch_load_b32 v63, off, s32 offset:8
; GFX11-TRUE16-NEXT:    scratch_load_b32 v62, off, s32 offset:12
; GFX11-TRUE16-NEXT:    scratch_load_b32 v61, off, s32 offset:16
; GFX11-TRUE16-NEXT:    scratch_load_b32 v60, off, s32 offset:20
; GFX11-TRUE16-NEXT:    scratch_load_b32 v59, off, s32 offset:24
; GFX11-TRUE16-NEXT:    scratch_load_b32 v58, off, s32 offset:28
; GFX11-TRUE16-NEXT:    scratch_load_b32 v57, off, s32 offset:32
; GFX11-TRUE16-NEXT:    scratch_load_b32 v56, off, s32 offset:36
; GFX11-TRUE16-NEXT:    scratch_load_b32 v47, off, s32 offset:40
; GFX11-TRUE16-NEXT:    scratch_load_b32 v46, off, s32 offset:44
; GFX11-TRUE16-NEXT:    scratch_load_b32 v45, off, s32 offset:48
; GFX11-TRUE16-NEXT:    scratch_load_b32 v44, off, s32 offset:52
; GFX11-TRUE16-NEXT:    scratch_load_b32 v43, off, s32 offset:56
; GFX11-TRUE16-NEXT:    scratch_load_b32 v42, off, s32 offset:60
; GFX11-TRUE16-NEXT:    scratch_load_b32 v41, off, s32 offset:64
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 offset:68
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, v32 :: v_dual_mov_b32 v1, v33
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, v34 :: v_dual_mov_b32 v3, v35
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, v36 :: v_dual_mov_b32 v5, v37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, v38 :: v_dual_mov_b32 v7, v39
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v64f16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB100_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v69, 0x40c00000, v69 :: v_dual_lshlrev_b32 v70, 16, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 16, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v83, 16, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v10
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v87, 16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v87, 0x40c00000, v87
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v16, v16, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 16, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v17
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v11
; GFX11-FAKE16-NEXT:    v_add3_u32 v17, v17, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v17, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v34, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v37, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v37, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v33, v17, 0x7060302
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v32, v16, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v34, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v37, v38, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v39, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v37, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v36, v39, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v39
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v34, v18, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v38, v37, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v37
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v48, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v38, v48, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v49, 0x40c00000, v20 :: v_dual_add_f32 v48, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v38, v39, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v39
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v49, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v48, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v50, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v35, v19, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v49, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v49
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v49, 0x40c00000, v22 :: v_dual_lshlrev_b32 v22, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v36, v20, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v39, v48, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v48
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v49, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v51, 0x40c00000, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v50, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v37, v21, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v39, v50, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v50
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v50, 0x40c00000, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v48, v49, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v49
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v51, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v50, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v39, v48, v51, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v51
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v52, 0x40c00000, v23 :: v_dual_add_f32 v51, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v39, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v39, v49, v50, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v50
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v51, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v53, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v38, v22, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v39, v23, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v49, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v52, 0x40c00000, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v50, v51, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v51
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v50, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v54, 0x40c00000, v25 :: v_dual_add_f32 v53, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v49, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v51, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v55, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v48, v24, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v49, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v54, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v49, v25, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v51, v54, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v54
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v54, 0x40c00000, v27
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v50, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v52, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v55, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v54, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v50, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v51, v52, v55, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v55
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v64, 0x40c00000, v27 :: v_dual_add_f32 v55, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v51, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v51, v53, v54, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v54
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-FAKE16-NEXT:    v_bfe_u32 v54, v55, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v65, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v50, v26, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v51, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v64, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v51, v27, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v53, v64, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v64
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v64, 0x40c00000, v29
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v52, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v54, v55, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v55
; GFX11-FAKE16-NEXT:    v_bfe_u32 v54, v65, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v30
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v64, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v52, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v54, v65, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v65
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v66, 0x40c00000, v29 :: v_dual_add_f32 v65, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v53, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v55, v64, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v64
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v65, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v67, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v52, v28, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v53, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v66, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v53, v29, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v54, v55, v66, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v66
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v66, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v54, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v54, v64, v65, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v65
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v67, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v65, v66, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v54, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v55, v64, v67, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v67
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v68, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v67, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v54, v30, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v55, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v55, v65, v66, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v66
; GFX11-FAKE16-NEXT:    v_bfe_u32 v65, v68, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v67, 0x40c00000, v67
; GFX11-FAKE16-NEXT:    v_bfe_u32 v66, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v55, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v64, v65, v68, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, 0x400000, v68
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v67, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v55, v31, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v64, v65, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v65, v66, v0, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v65, v66 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-FAKE16-NEXT:    v_add3_u32 v65, v68, v67, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v67
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v64, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v65, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v66, v68, v1, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v69, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v66, v67, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v66, v68, v69, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v69
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v65, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v66, v67, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v67, v68, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v68, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v67, v68 :: v_dual_and_b32 v3, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v70, 0x40c00000, v70 :: v_dual_add_f32 v3, 0x40c00000, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v66, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v69, v70, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v68, 0x400000, v70
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v70, v70
; GFX11-FAKE16-NEXT:    v_add3_u32 v67, v69, v70, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v69, v3, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v67, v67, v68, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v68, v69, v3, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v68, v69 :: v_dual_and_b32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v71, 0x40c00000, v71 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v67, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v70, v71, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, 0x400000, v71
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v71, v71
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v68, v70, v71, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v70, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v68, v68, v69, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v69, v70, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v70, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v69, v70 :: v_dual_and_b32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v80, 0x40c00000, v80 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v68, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v80, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v70, 0x400000, v80
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-FAKE16-NEXT:    v_add3_u32 v69, v71, v80, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v69, v70, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v70, v71, v5, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v70, v71 :: v_dual_and_b32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v81, 0x40c00000, v81 :: v_dual_add_f32 v6, 0x40c00000, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v69, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v80, v81, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v81
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v70, v80, v81, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v80, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v70, v70, v71, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v71, v80, v6, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v80, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v71, v80 :: v_dual_and_b32 v7, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v82, 0x40c00000, v82 :: v_dual_add_f32 v7, 0x40c00000, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v70, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v81, v82, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v80, 0x400000, v82
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-FAKE16-NEXT:    v_add3_u32 v71, v81, v82, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v81, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v71, v71, v80, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v80, v81, v7, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v81, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v80, v81 :: v_dual_and_b32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v83, 0x40c00000, v83 :: v_dual_add_f32 v8, 0x40c00000, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v71, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v83, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v81, 0x400000, v83
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v80, v82, v83, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v80, v80, v81, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v81, v82, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v81, v82 :: v_dual_and_b32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v84, 0x40c00000, v84 :: v_dual_add_f32 v9, 0x40c00000, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v80, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v84, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, 0x400000, v84
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-FAKE16-NEXT:    v_add3_u32 v81, v83, v84, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v9, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v81, v81, v82, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v82, v83, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v82, v83 :: v_dual_and_b32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v85, 0x40c00000, v85 :: v_dual_add_f32 v10, 0x40c00000, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v81, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v84, v85, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, 0x400000, v85
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v82, v84, v85, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v84, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v82, v82, v83, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v83, v84, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v83, v84 :: v_dual_and_b32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v86, 0x40c00000, v86 :: v_dual_add_f32 v11, 0x40c00000, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v82, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v85, v86, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, 0x400000, v86
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v96, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_add3_u32 v83, v85, v86, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v85, v11, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v83, v83, v84 :: v_dual_lshlrev_b32 v84, 16, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v86, v87, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v97, 0x400000, v87
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v84, 0x40c00000, v84
; GFX11-FAKE16-NEXT:    v_add3_u32 v86, v86, v87, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v98, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v84, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v85, v85, v11, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v86, v86, v97, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v97, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_add3_u32 v87, v98, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v98, v99, v84, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v99, 16, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v100, 0x400000, v84
; GFX11-FAKE16-NEXT:    v_bfe_u32 v101, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v87, v97 :: v_dual_add_f32 v87, 0x40c00000, v99
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v97, v101, v13, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v86, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v87, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v84, v98, v100, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v101, 0x400000, v87
; GFX11-FAKE16-NEXT:    v_bfe_u32 v102, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v99, v99, v87, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v100, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v98, 0x40c00000, v98 :: v_dual_cndmask_b32 v87, v99, v101
; GFX11-FAKE16-NEXT:    v_add3_u32 v101, v102, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v102, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v103, v98, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v112, 0x400000, v98
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v101, v102 :: v_dual_add_f32 v15, 0x40c00000, v15
; GFX11-FAKE16-NEXT:    v_add3_u32 v103, v103, v98, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v87, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v113, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v98, v103, v112, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v99, v99, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v99, v113, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v98, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v97, v100, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v84, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v85, v96, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v83, 0x7060302
; GFX11-FAKE16-NEXT:  .LBB100_2: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define inreg <64 x half> @bitcast_v64bf16_to_v64f16_scalar(<64 x bfloat> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64bf16_to_v64f16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:688 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; SI-NEXT:    v_readfirstlane_b32 s92, v4
; SI-NEXT:    s_and_b32 s30, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v1
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_readfirstlane_b32 s78, v7
; SI-NEXT:    s_and_b32 s93, s92, 0xffff0000
; SI-NEXT:    s_and_b32 s34, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    s_and_b32 s79, s78, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s93
; SI-NEXT:    v_readfirstlane_b32 s46, v15
; SI-NEXT:    s_and_b32 s77, s76, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s79
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s77
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s47
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s45
; SI-NEXT:    s_and_b32 s14, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s15, s24, 16
; SI-NEXT:    s_and_b32 s24, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s17
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s23
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s15
; SI-NEXT:    s_lshl_b32 s11, s26, 16
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    v_readfirstlane_b32 s72, v10
; SI-NEXT:    s_and_b32 s73, s72, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    v_readfirstlane_b32 s60, v12
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    v_readfirstlane_b32 s56, v14
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_readfirstlane_b32 s58, v13
; SI-NEXT:    v_readfirstlane_b32 s62, v11
; SI-NEXT:    v_readfirstlane_b32 s74, v9
; SI-NEXT:    v_readfirstlane_b32 s88, v6
; SI-NEXT:    v_readfirstlane_b32 s90, v5
; SI-NEXT:    v_readfirstlane_b32 s94, v3
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s29, 16
; SI-NEXT:    s_lshl_b32 s7, s28, 16
; SI-NEXT:    s_and_b32 s8, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s9, s27, 16
; SI-NEXT:    s_and_b32 s10, s26, 0xffff0000
; SI-NEXT:    s_and_b32 s12, s25, 0xffff0000
; SI-NEXT:    s_lshl_b32 s13, s25, 16
; SI-NEXT:    s_and_b32 s25, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s26, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s27, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s28, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s29, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s41, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s11
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_and_b32 s63, s62, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_and_b32 s75, s74, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_and_b32 s89, s88, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_and_b32 s91, s90, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_and_b32 s95, s94, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_and_b32 s36, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v18
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cmp_lg_u32 vcc_lo, 0
; SI-NEXT:    v_mul_f32_e64 v39, 1.0, s41
; SI-NEXT:    v_mul_f32_e64 v44, 1.0, s40
; SI-NEXT:    v_mul_f32_e64 v47, 1.0, s29
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s28
; SI-NEXT:    v_mul_f32_e64 v53, 1.0, s27
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s26
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s25
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s24
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v62, 1.0, s14
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s12
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s10
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s8
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s4
; SI-NEXT:    v_mul_f32_e64 v51, 1.0, s36
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s34
; SI-NEXT:    v_mul_f32_e64 v38, 1.0, s30
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s95
; SI-NEXT:    v_mul_f32_e64 v32, 1.0, s91
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s89
; SI-NEXT:    v_mul_f32_e64 v42, 1.0, s75
; SI-NEXT:    v_mul_f32_e64 v55, 1.0, s73
; SI-NEXT:    v_mul_f32_e64 v56, 1.0, s63
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s61
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s31
; SI-NEXT:    v_mul_f32_e64 v37, 1.0, s59
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s19
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s21
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s13
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s9
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s5
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s35
; SI-NEXT:    v_mul_f32_e64 v35, 1.0, s94
; SI-NEXT:    v_mul_f32_e64 v34, 1.0, s90
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s78
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s74
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s62
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s58
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s46
; SI-NEXT:    v_mul_f32_e64 v33, 1.0, s42
; SI-NEXT:    v_mul_f32_e64 v52, 1.0, s16
; SI-NEXT:    v_mul_f32_e64 v50, 1.0, s18
; SI-NEXT:    v_mul_f32_e64 v41, 1.0, s20
; SI-NEXT:    v_mul_f32_e64 v40, 1.0, s22
; SI-NEXT:    v_mul_f32_e64 v45, 1.0, s7
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s37
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s88
; SI-NEXT:    v_mul_f32_e64 v61, 1.0, s76
; SI-NEXT:    v_mul_f32_e64 v60, 1.0, s57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s92
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s72
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:520 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:524 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s60
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:528 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:532 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s56
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:536 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:540 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:552 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:556 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:560 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:564 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v51, off, s[0:3], s32 offset:568 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:572 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:576 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:580 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:584 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:588 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:592 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:596 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:600 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:604 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:608 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:612 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:616 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:620 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:624 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:628 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:632 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB101_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_mov_b32_e32 v43, v44
; SI-NEXT:    v_mov_b32_e32 v60, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v3
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v39
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v18
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v21
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v29
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v42
; SI-NEXT:    buffer_load_dword v48, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v49, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v53
; SI-NEXT:    v_mov_b32_e32 v24, v6
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v60
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v38
; SI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    v_lshr_b64 v[1:2], v[1:2], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[7:8], 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:636 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:640 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:644 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[4:5], 16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v53
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:648 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:652 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:656 ; 4-byte Folded Spill
; SI-NEXT:    v_lshr_b64 v[3:4], v[13:14], 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v62
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[3:4], v[19:20], 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v1, v52
; SI-NEXT:    v_lshr_b64 v[13:14], v[16:17], 16
; SI-NEXT:    v_mov_b32_e32 v52, v15
; SI-NEXT:    v_lshr_b64 v[15:16], v[22:23], 16
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v51
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:660 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:664 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:668 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[45:46], v[35:36], 16
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v32
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v16, v37
; SI-NEXT:    v_mov_b32_e32 v0, v50
; SI-NEXT:    v_mov_b32_e32 v5, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v5
; SI-NEXT:    v_mov_b32_e32 v14, v61
; SI-NEXT:    v_mov_b32_e32 v23, v43
; SI-NEXT:    v_lshr_b64 v[10:11], v[48:49], 16
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v26
; SI-NEXT:    v_mov_b32_e32 v11, v26
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v8
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v19, v8
; SI-NEXT:    v_lshr_b64 v[7:8], v[28:29], 16
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v16
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_mov_b32_e32 v4, v56
; SI-NEXT:    v_mov_b32_e32 v3, v55
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v4
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v20
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v22
; SI-NEXT:    v_lshr_b64 v[58:59], v[34:35], 16
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v55
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    v_lshr_b64 v[55:56], v[27:28], 16
; SI-NEXT:    v_mov_b32_e32 v56, v4
; SI-NEXT:    s_waitcnt vmcnt(5) expcnt(4)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v8
; SI-NEXT:    v_lshr_b64 v[36:37], v[25:26], 16
; SI-NEXT:    v_mov_b32_e32 v37, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v39
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:672 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:676 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v47
; SI-NEXT:    v_mov_b32_e32 v47, v13
; SI-NEXT:    v_mov_b32_e32 v46, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v19
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:680 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:684 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, v41
; SI-NEXT:    v_mov_b32_e32 v1, v42
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v5
; SI-NEXT:    v_lshr_b64 v[49:50], v[31:32], 16
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v51
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v15
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v2, v8
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v22
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v62
; SI-NEXT:    v_mov_b32_e32 v17, v16
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v19, v24
; SI-NEXT:    v_lshr_b64 v[29:30], v[30:31], 16
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v21
; SI-NEXT:    v_lshr_b64 v[31:32], v[33:34], 16
; SI-NEXT:    v_mov_b32_e32 v26, v30
; SI-NEXT:    v_mov_b32_e32 v27, v31
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v21
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v60
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v2
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v52
; SI-NEXT:    v_mov_b32_e32 v53, v29
; SI-NEXT:    v_mov_b32_e32 v52, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v38
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[34:35], v[35:36], 16
; SI-NEXT:    v_mov_b32_e32 v51, v4
; SI-NEXT:    v_mov_b32_e32 v50, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:636 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:640 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:644 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v8, v4
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:648 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:652 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:656 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mov_b32_e32 v5, v10
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[9:10], v[50:51], 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[9:10], v[46:47], 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v22, v4
; SI-NEXT:    v_mov_b32_e32 v21, v3
; SI-NEXT:    v_mov_b32_e32 v3, v47
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v3, v45
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v22, v7
; SI-NEXT:    v_mov_b32_e32 v47, v36
; SI-NEXT:    v_lshr_b64 v[35:36], v[0:1], 16
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:660 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:664 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:668 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v3, v58
; SI-NEXT:    v_lshr_b64 v[57:58], v[57:58], 16
; SI-NEXT:    v_lshr_b64 v[58:59], v[48:49], 16
; SI-NEXT:    v_mov_b32_e32 v59, v53
; SI-NEXT:    v_mov_b32_e32 v48, v0
; SI-NEXT:    v_lshr_b64 v[0:1], v[40:41], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[42:43], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mov_b32_e32 v21, v10
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[9:10], v[44:45], 16
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[0:1], v[32:33], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v44, v23
; SI-NEXT:    v_mov_b32_e32 v23, v38
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v3, v49
; SI-NEXT:    v_lshr_b64 v[49:50], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[6:7], v[54:55], 16
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:672 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:676 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[53:54], v[52:53], 16
; SI-NEXT:    v_lshr_b64 v[51:52], v[26:27], 16
; SI-NEXT:    v_mov_b32_e32 v54, v27
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshr_b64 v[0:1], v[45:46], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[17:18], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mov_b32_e32 v52, v6
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshr_b64 v[26:27], v[6:7], 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:680 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:684 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[28:29], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[19:20], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[14:15], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[12:13], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[24:25], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[60:61], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[30:31], 16
; SI-NEXT:    v_mov_b32_e32 v46, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v9, v28
; SI-NEXT:    v_mov_b32_e32 v3, v32
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mov_b32_e32 v50, v6
; SI-NEXT:    v_lshr_b64 v[10:11], v[6:7], 16
; SI-NEXT:    v_mov_b32_e32 v11, v40
; SI-NEXT:    v_lshr_b64 v[40:41], v[38:39], 16
; SI-NEXT:    s_cbranch_execnz .LBB101_3
; SI-NEXT:  .LBB101_2: ; %cmp.true
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v45
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v23
; SI-NEXT:    s_waitcnt vmcnt(3) expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v50
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v52
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v19
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v60
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v3
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v24
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v30
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v48
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v42
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_add_f32_e32 v34, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v45, 0x40c00000, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v37
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v56
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v36, 0x40c00000, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:632 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v38, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:628 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v52, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:624 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:620 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v56, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:616 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v58, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:612 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v40, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:608 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v60, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:604 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v15
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:600 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v54, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:596 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v44
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v17
; SI-NEXT:    v_lshr_b64 v[43:44], v[34:35], 16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:592 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:588 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v24
; SI-NEXT:    v_lshr_b64 v[47:48], v[16:17], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v30
; SI-NEXT:    v_lshr_b64 v[15:16], v[0:1], 16
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:584 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v62, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v62
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:580 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v27
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v2
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:576 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v20
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:572 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[28:29], v[36:37], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v5
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v21
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v6
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:568 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v7
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v6
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:564 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v8
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v7
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v9
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v8
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:560 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v10
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v9
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v11
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v10
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:556 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v11
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:552 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v13
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v12
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v14
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v13
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v33
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v14
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v35
; SI-NEXT:    v_lshr_b64 v[34:35], v[22:23], 16
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v12
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v10
; SI-NEXT:    v_lshr_b64 v[10:11], v[49:50], 16
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v8
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v7
; SI-NEXT:    v_lshr_b64 v[7:8], v[40:41], 16
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v14
; SI-NEXT:    v_lshr_b64 v[13:14], v[38:39], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v46
; SI-NEXT:    v_lshr_b64 v[45:46], v[45:46], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v23
; SI-NEXT:    v_lshr_b64 v[22:23], v[25:26], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v37
; SI-NEXT:    v_lshr_b64 v[36:37], v[52:53], 16
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v6
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v39
; SI-NEXT:    v_lshr_b64 v[38:39], v[56:57], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v50
; SI-NEXT:    v_lshr_b64 v[50:51], v[18:19], 16
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v62
; SI-NEXT:    v_lshr_b64 v[17:18], v[58:59], 16
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v53
; SI-NEXT:    v_lshr_b64 v[52:53], v[54:55], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v19
; SI-NEXT:    v_lshr_b64 v[19:20], v[60:61], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v57
; SI-NEXT:    v_lshr_b64 v[57:58], v[35:36], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v59
; SI-NEXT:    v_lshr_b64 v[58:59], v[9:10], 16
; SI-NEXT:    v_mov_b32_e32 v59, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v41
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v61
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v26
; SI-NEXT:    v_mov_b32_e32 v26, v47
; SI-NEXT:    v_lshr_b64 v[46:47], v[49:50], 16
; SI-NEXT:    v_mov_b32_e32 v47, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v55
; SI-NEXT:    v_mov_b32_e32 v55, v34
; SI-NEXT:    v_mov_b32_e32 v54, v33
; SI-NEXT:    v_lshr_b64 v[34:35], v[27:28], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v24
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v32
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[31:32], 16
; SI-NEXT:    v_mov_b32_e32 v2, v4
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v24, v7
; SI-NEXT:    v_mov_b32_e32 v23, v6
; SI-NEXT:    v_lshr_b64 v[6:7], v[3:4], 16
; SI-NEXT:    v_mov_b32_e32 v4, v52
; SI-NEXT:    v_mov_b32_e32 v8, v6
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v2, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[5:6], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[51:52], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[21:22], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v19
; SI-NEXT:    v_lshr_b64 v[51:52], v[42:43], 16
; SI-NEXT:    v_mov_b32_e32 v5, v22
; SI-NEXT:    v_mov_b32_e32 v21, v38
; SI-NEXT:    v_mov_b32_e32 v22, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshr_b64 v[0:1], v[18:19], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshr_b64 v[0:1], v[23:24], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v17
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshr_b64 v[0:1], v[16:17], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[37:38], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v50
; SI-NEXT:    v_lshr_b64 v[49:50], v[12:13], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, v36
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v1, v10
; SI-NEXT:    v_mov_b32_e32 v10, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshr_b64 v[0:1], v[54:55], 16
; SI-NEXT:    v_lshr_b64 v[53:54], v[44:45], 16
; SI-NEXT:    v_mov_b32_e32 v54, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:  .LBB101_3: ; %end
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v4
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v5
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v26
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v8
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v58
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v12
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v21
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v4
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v14
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v57
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v25
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v2, v2, v4
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v25
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v35
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v25
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_or_b32_e32 v6, v6, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v17
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v7, v7, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v8, v10, v8
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v39
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v39
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v16
; SI-NEXT:    v_or_b32_e32 v12, v12, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v25
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v40
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v14, v14, v16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v25
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v46
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v25
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v17, v18, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v18, v18, v20
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v25
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v20, v20, v22
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v25
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v49
; SI-NEXT:    v_or_b32_e32 v22, v22, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v34
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v47
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v35
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v29
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v55
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v35
; SI-NEXT:    v_or_b32_e32 v27, v27, v29
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v53
; SI-NEXT:    v_or_b32_e32 v28, v29, v28
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v59
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v35
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v51
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v54
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:688 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB101_4:
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v25, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    v_mov_b32_e32 v11, v40
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:528 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:532 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:520 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:524 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v48, v41
; SI-NEXT:    v_mov_b32_e32 v14, v61
; SI-NEXT:    v_mov_b32_e32 v19, v6
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:536 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:540 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:544 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:548 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mov_b32_e32 v17, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v60, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:512 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:516 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v9, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mov_b32_e32 v3, v0
; SI-NEXT:    s_branch .LBB101_2
;
; VI-LABEL: bitcast_v64bf16_to_v64f16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v35, s30, 0
; VI-NEXT:    v_writelane_b32 v35, s31, 1
; VI-NEXT:    v_writelane_b32 v35, s36, 2
; VI-NEXT:    v_writelane_b32 v35, s37, 3
; VI-NEXT:    v_writelane_b32 v35, s38, 4
; VI-NEXT:    v_writelane_b32 v35, s39, 5
; VI-NEXT:    v_writelane_b32 v35, s48, 6
; VI-NEXT:    v_writelane_b32 v35, s49, 7
; VI-NEXT:    v_writelane_b32 v35, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v35, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB101_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB101_4
; VI-NEXT:  .LBB101_2: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x40c00000
; VI-NEXT:    s_lshl_b32 s4, s48, 16
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_and_b32 s4, s48, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v16, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    s_lshl_b32 s5, s28, 16
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    s_and_b32 s5, s28, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    v_add_f32_e32 v1, s5, v32
; VI-NEXT:    v_bfe_u32 v2, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s5, s30, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; VI-NEXT:    v_add_f32_e32 v2, s5, v32
; VI-NEXT:    v_bfe_u32 v3, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    s_and_b32 s5, s30, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; VI-NEXT:    v_add_f32_e32 v3, s5, v32
; VI-NEXT:    v_bfe_u32 v4, v3, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; VI-NEXT:    s_and_b32 s5, s26, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_lshrrev_b64 v[12:13], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s26, 16
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_lshrrev_b64 v[14:15], 16, v[2:3]
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s24, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[10:11], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s24, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s22, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[8:9], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s22, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s20, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s20, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s18, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[4:5], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s18, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s16, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[2:3], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s16, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; VI-NEXT:    v_bfe_u32 v3, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v0
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s17, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[0:1], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v1, s5, v32
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_lshl_b32 s4, s17, 16
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_and_b32 s5, s19, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_add_f32_e32 v3, s5, v32
; VI-NEXT:    v_cndmask_b32_e32 v18, v5, v7, vcc
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_lshl_b32 s4, s19, 16
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_and_b32 s5, s21, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_add_f32_e32 v5, s5, v32
; VI-NEXT:    v_mov_b32_e32 v1, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v7, v9, vcc
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_lshl_b32 s4, s21, 16
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v3
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_and_b32 s5, s23, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_add_f32_e32 v7, s5, v32
; VI-NEXT:    v_mov_b32_e32 v3, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v9, v11, vcc
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_lshl_b32 s4, s23, 16
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v5
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_and_b32 s5, s25, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_add_f32_e32 v9, s5, v32
; VI-NEXT:    v_mov_b32_e32 v5, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v11, v13, vcc
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_lshl_b32 s4, s25, 16
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v7
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_and_b32 s5, s27, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_add_f32_e32 v11, s5, v32
; VI-NEXT:    v_mov_b32_e32 v7, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v13, v15, vcc
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_lshl_b32 s4, s27, 16
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v9
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_and_b32 s5, s29, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v9, v18
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_add_f32_e32 v13, s5, v32
; VI-NEXT:    v_cndmask_b32_e32 v18, v15, v18, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v11
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_lshl_b32 s4, s29, 16
; VI-NEXT:    v_mov_b32_e32 v11, v18
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v15, v18, vcc
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_bfe_u32 v18, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v15
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v13
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    s_and_b32 s5, s31, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v15, s5, v32
; VI-NEXT:    v_mov_b32_e32 v13, v18
; VI-NEXT:    v_bfe_u32 v18, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v15
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    s_lshl_b32 s4, s31, 16
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v18, v19, vcc
; VI-NEXT:    v_add_f32_e32 v18, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v18
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v15
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    s_and_b32 s5, s50, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v15, v18
; VI-NEXT:    v_add_f32_e32 v18, s5, v32
; VI-NEXT:    v_bfe_u32 v19, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v18
; VI-NEXT:    s_lshl_b32 s4, s50, 16
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_add_f32_e32 v18, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc
; VI-NEXT:    v_bfe_u32 v20, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v20, vcc, v20, v18
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x7fff, v20
; VI-NEXT:    v_lshrrev_b64 v[28:29], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v18, v20, v21, vcc
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s46, 16
; VI-NEXT:    v_lshrrev_b64 v[30:31], 16, v[18:19]
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[26:27], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[24:25], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[22:23], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s38, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[20:21], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s36, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v19, vcc
; VI-NEXT:    v_bfe_u32 v19, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v16
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v19, v21, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshrrev_b64 v[16:17], 16, v[16:17]
; VI-NEXT:    s_and_b32 s4, s37, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    s_lshl_b32 s4, s37, 16
; VI-NEXT:    v_cndmask_b32_e32 v17, v19, v21, vcc
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    s_and_b32 s4, s39, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v21, v23, vcc
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    s_lshl_b32 s4, s39, 16
; VI-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_mov_b32_e32 v17, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v23, v25, vcc
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    s_lshl_b32 s4, s41, 16
; VI-NEXT:    v_cndmask_b32_e32 v21, v23, v25, vcc
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_mov_b32_e32 v19, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v25, v27, vcc
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    s_lshl_b32 s4, s43, 16
; VI-NEXT:    v_cndmask_b32_e32 v23, v25, v27, vcc
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v21
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_mov_b32_e32 v21, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v27, v29, vcc
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    s_lshl_b32 s4, s45, 16
; VI-NEXT:    v_cndmask_b32_e32 v25, v27, v29, vcc
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_mov_b32_e32 v23, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v29, v31, vcc
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    s_lshl_b32 s4, s47, 16
; VI-NEXT:    v_cndmask_b32_e32 v27, v29, v31, vcc
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    s_and_b32 s4, s49, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v25, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v31, v33, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    v_mov_b32_e32 v27, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    s_lshl_b32 s4, s49, 16
; VI-NEXT:    v_cndmask_b32_e32 v29, v31, v33, vcc
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    s_and_b32 s4, s51, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_mov_b32_e32 v29, v33
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    s_lshl_b32 s4, s51, 16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v32, v31, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_add_u32_e32 v32, vcc, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x7fff, v32
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v32, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; VI-NEXT:    v_lshrrev_b64 v[31:32], 16, v[31:32]
; VI-NEXT:    s_branch .LBB101_5
; VI-NEXT:  .LBB101_3:
; VI-NEXT:    s_branch .LBB101_2
; VI-NEXT:  .LBB101_4:
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:  .LBB101_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v35, 9
; VI-NEXT:    v_readlane_b32 s50, v35, 8
; VI-NEXT:    v_readlane_b32 s49, v35, 7
; VI-NEXT:    v_readlane_b32 s48, v35, 6
; VI-NEXT:    v_readlane_b32 s39, v35, 5
; VI-NEXT:    v_readlane_b32 s38, v35, 4
; VI-NEXT:    v_readlane_b32 s37, v35, 3
; VI-NEXT:    v_readlane_b32 s36, v35, 2
; VI-NEXT:    v_readlane_b32 s31, v35, 1
; VI-NEXT:    v_readlane_b32 s30, v35, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v64f16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v36, s30, 0
; GFX9-NEXT:    v_writelane_b32 v36, s31, 1
; GFX9-NEXT:    v_writelane_b32 v36, s36, 2
; GFX9-NEXT:    v_writelane_b32 v36, s37, 3
; GFX9-NEXT:    v_writelane_b32 v36, s38, 4
; GFX9-NEXT:    v_writelane_b32 v36, s39, 5
; GFX9-NEXT:    v_writelane_b32 v36, s48, 6
; GFX9-NEXT:    v_writelane_b32 v36, s49, 7
; GFX9-NEXT:    v_writelane_b32 v36, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v36, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB101_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB101_4
; GFX9-NEXT:  .LBB101_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x40c00000
; GFX9-NEXT:    s_and_b32 s4, s50, 0xffff0000
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    s_and_b32 s5, s30, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_add_f32_e32 v0, s5, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v1, v2, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s5, s30, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s5, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    s_and_b32 s5, s31, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_add_f32_e32 v2, s5, v16
; GFX9-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v3, v3, v2
; GFX9-NEXT:    v_add_u32_e32 v3, 0x7fff, v3
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    s_lshl_b32 s5, s31, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT:    v_add_f32_e32 v3, s5, v16
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    v_mov_b32_e32 v32, 0xffff
; GFX9-NEXT:    s_and_b32 s4, s29, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_lshl_or_b32 v14, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX9-NEXT:    v_and_b32_sdwa v3, v32, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_lshl_or_b32 v15, v2, 16, v3
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s29, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s28, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v13, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s28, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s27, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v12, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s27, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s26, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v11, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s26, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s25, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v10, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s25, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s24, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v9, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s24, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s23, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v8, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s23, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s22, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v7, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s22, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s21, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v6, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s21, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s20, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v5, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s20, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s19, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v4, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s19, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s18, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v3, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s18, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v18, vcc
; GFX9-NEXT:    s_and_b32 s4, s17, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v2, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s17, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v18, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v1
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v18, v19, vcc
; GFX9-NEXT:    s_and_b32 s4, s16, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v1, v32, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v1, v0, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v0
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s16, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    s_lshl_b32 s4, s50, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshl_or_b32 v0, v0, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    s_and_b32 s4, s51, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_add_f32_e32 v19, s4, v16
; GFX9-NEXT:    v_bfe_u32 v20, v19, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v20, v20, v19
; GFX9-NEXT:    v_add_u32_e32 v20, 0x7fff, v20
; GFX9-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    s_lshl_b32 s4, s51, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v20, v21, vcc
; GFX9-NEXT:    v_add_f32_e32 v20, s4, v16
; GFX9-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v21, v21, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s49, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v21, 0x7fff, v21
; GFX9-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_lshl_or_b32 v30, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v21, v22, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; GFX9-NEXT:    v_and_b32_sdwa v20, v32, v20 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_lshl_or_b32 v31, v19, 16, v20
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s49, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s48, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v29, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s48, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s47, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v28, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s47, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s46, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v27, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s46, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s45, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v26, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s45, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s44, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v25, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s44, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s43, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v24, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s43, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s42, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v23, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s42, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s41, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v22, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s41, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s40, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v21, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s40, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s39, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v20, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s39, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v33, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s38, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v19, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s38, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v33, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v18
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v18, v32, v18 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s37, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v18, v17, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v17
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s37, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; GFX9-NEXT:    v_and_b32_sdwa v33, v32, v33 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    s_and_b32 s4, s36, 0xffff0000
; GFX9-NEXT:    v_lshl_or_b32 v17, v17, 16, v33
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    s_lshl_b32 s4, s36, 16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v16, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v16, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v34, v35, vcc
; GFX9-NEXT:    v_and_b32_sdwa v16, v32, v16 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1
; GFX9-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; GFX9-NEXT:    v_lshl_or_b32 v16, v32, 16, v16
; GFX9-NEXT:    s_branch .LBB101_5
; GFX9-NEXT:  .LBB101_3:
; GFX9-NEXT:    s_branch .LBB101_2
; GFX9-NEXT:  .LBB101_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:  .LBB101_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v36, 9
; GFX9-NEXT:    v_readlane_b32 s50, v36, 8
; GFX9-NEXT:    v_readlane_b32 s49, v36, 7
; GFX9-NEXT:    v_readlane_b32 s48, v36, 6
; GFX9-NEXT:    v_readlane_b32 s39, v36, 5
; GFX9-NEXT:    v_readlane_b32 s38, v36, 4
; GFX9-NEXT:    v_readlane_b32 s37, v36, 3
; GFX9-NEXT:    v_readlane_b32 s36, v36, 2
; GFX9-NEXT:    v_readlane_b32 s31, v36, 1
; GFX9-NEXT:    v_readlane_b32 s30, v36, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v64f16_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s0
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s37, s29
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-TRUE16-NEXT:    s_mov_b32 s36, s28
; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s3
; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s2
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s1
; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB101_3
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB101_4
; GFX11-TRUE16-NEXT:  .LBB101_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s37, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    s_and_b32 s2, s38, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s25, 0xffff0000
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v2, v6 :: v_dual_add_nc_u32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v85, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v6, v2 :: v_dual_add_nc_u32 v2, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s27, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v2, v3 :: v_dual_add_nc_u32 v3, v4, v0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v87, 0x400000, v85
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v2, v3 :: v_dual_add_nc_u32 v3, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v2, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v97, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v5, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v98, v97, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v98, v98, v97
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.h, v36.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.h, v35.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v98, 0x7fff, v98
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.h, v37.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.h, v34.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.h, v33.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.h, v32.l
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.h, v38.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.h, v39.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.h, v48.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.h, v49.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.h, v50.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.h, v51.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s50, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.h, v52.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.h, v53.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s12, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s13, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.h, v54.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.h, v55.l
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s14, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v0, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s14, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v4, v5 :: v_dual_add_nc_u32 v4, v6, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v3
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s15, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v4, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v2, v4
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s16, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v7, v9
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v5, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s17, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v4, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v10, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v4
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v5, v9 :: v_dual_add_nc_u32 v8, 0x7fff, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v4, v7
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v5
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s18, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v10
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v5, v8, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v6, v8 :: v_dual_add_nc_u32 v6, v9, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s19, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v10
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v6, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v12, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v6
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v9, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v7, v11 :: v_dual_add_nc_u32 v10, 0x7fff, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v6, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v7
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s20, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v7, v10, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.h, v70.l
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v8, v10 :: v_dual_add_nc_u32 v8, v11, v13
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s21, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v8, v11, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v14, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v9, v13 :: v_dual_add_nc_u32 v12, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v8, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v9
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s22, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v12, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.h, v80.l
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v10, v12 :: v_dual_add_nc_u32 v10, v13, v15
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s23, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v82, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v10, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v82, v82, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v10
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v11, v11, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v82
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v82, v10, v13
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v84, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s24, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v82
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v14, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v84, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s25, 16
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.h, v83.l
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v12, v12, v14 :: v_dual_add_nc_u32 v13, v15, v84
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v82, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v85, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.h, v81.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v12
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v14, v82
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v15, v85
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v84, 0x400000, v82
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.h, v86.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v12, v13, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-TRUE16-NEXT:    v_bfe_u32 v96, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v14, v84, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s26, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v84, v96, v13
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v12
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v15, v87, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v87, v82, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v84
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v99, 0x400000, v82
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v15
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v84, v87, v82
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v87, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s27, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.h, v85.l
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v14, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v87, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v84, 0x7fff, v84
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v100, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, v15, v87
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v101, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v82, v84, v99, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v99, 0x400000, v97
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v97, v97
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v84, v100, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v100, 0x400000, v87
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v82
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v97, v98, v99, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v84, 0x7fff, v84
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.h, v96.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.h, v71.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v97
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v15, v100, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.h, v69.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.h, v68.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.h, v67.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v14, 16, v15
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v84, v84, v101, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.h, v82.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.h, v66.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.h, v65.l
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.h, v64.l
; GFX11-TRUE16-NEXT:    v_lshrrev_b32_e32 v15, 16, v84
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.h, v87.l
; GFX11-TRUE16-NEXT:    s_branch .LBB101_5
; GFX11-TRUE16-NEXT:  .LBB101_3:
; GFX11-TRUE16-NEXT:    s_branch .LBB101_2
; GFX11-TRUE16-NEXT:  .LBB101_4:
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-TRUE16-NEXT:  .LBB101_5: ; %end
; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v64f16_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s12, s0
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s37, s29
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-FAKE16-NEXT:    s_mov_b32 s36, s28
; GFX11-FAKE16-NEXT:    s_mov_b32 s15, s3
; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s2
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s1
; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB101_3
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB101_4
; GFX11-FAKE16-NEXT:  .LBB101_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s37, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v3, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    s_and_b32 s2, s38, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v7, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v3, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v2, v6 :: v_dual_add_nc_u32 v3, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v16, 16, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v9, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v17, 16, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v1, v6, v2 :: v_dual_add_nc_u32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v3 :: v_dual_add_nc_u32 v3, v4, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v18, 16, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v2, v3 :: v_dual_add_nc_u32 v3, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v2, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v2
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v34, 16, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v5, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v20, 16, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v35, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v36, 0xffff, v36
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v21, 16, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v22, 16, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v23, 16, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v24, 16, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v25, 16, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v49, 0xffff, v49
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v26, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v27, 16, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v28, 16, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, v0, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v29, 16, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v3, v4, vcc_lo
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s50, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v30, 16, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v2, v4 :: v_dual_add_nc_u32 v3, 0x7fff, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v0, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_add_nc_u32 v0, v0, v1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s12, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v31, 16, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v5, v2
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v7, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s13, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v3
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v54, 0xffff, v54
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v7, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v1, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s14, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s14, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, v1, v6
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v2, v2, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v5, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v3, v4 :: v_dual_add_nc_u32 v4, 0x7fff, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s15, 16
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v1, 16, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v8, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v6
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v8 :: v_dual_add_nc_u32 v5, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v5, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v8, v9, vcc_lo
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s16, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v6
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v9, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s17, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v4, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff, v66
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v6
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v9
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v4, v7 :: v_dual_add_nc_u32 v7, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v7, v8 :: v_dual_and_b32 v67, 0xffff, v67
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v11, v5
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s18, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v7, v8 :: v_dual_add_nc_u32 v8, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v8, v9 :: v_dual_add_nc_u32 v8, v10, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s19, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v9, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v10, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v8
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v9, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s20, 16
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v5, v68, 16, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v9, v12 :: v_dual_and_b32 v68, 0xffff, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v12, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v3, v65, 16, v67
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v8, v8, v10, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v12, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s21, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, v11, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v65, 0xffff, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v11, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v80, 16, v9
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s22, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v11, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v12, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v9, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v81, 16, v11
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v12, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v14, v9
; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s23, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v12, v15 :: v_dual_and_b32 v80, 0xffff, v80
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v82, v82, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v12, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v1, v1, 16, v68
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v82
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v82, v83, v12
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s24, 16
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v13, v13, v15 :: v_dual_add_nc_u32 v14, 0x7fff, v82
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v83, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v84, 16, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, 0x400000, v83
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v12, v14, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v82, v83
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s25, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v86, v82, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v14, v85 :: v_dual_add_nc_u32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v86, v86, v82
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v14, v83, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v100, 0x400000, v83
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v15, v85, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v86
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v86, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s26, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v85, 0x400000, v82
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s27, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v102, 0x400000, v86
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v96, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s27, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v97, v82, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v15, v85, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v85, v86, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v96, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v98, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, v97, v82
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v103, 0x400000, v82
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, v99, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v85, v85, v86
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v97
; GFX11-FAKE16-NEXT:    v_bfe_u32 v101, v98, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v112, 0x400000, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, 0x7fff, v99
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v82, v97, v103 :: v_dual_add_nc_u32 v85, 0x7fff, v85
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v96, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v101, v101, v98
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v14, v83
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v82
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v96, v99, v112, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v101
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v101, 0x400000, v98
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v96
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v85, v85, v102, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff, v82
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v86, v97, v101, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v83, 16, v85
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v97, 16, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v86
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v14, v14, v100, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v86, 0xffff, v96
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v14
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v14, v83, 16, v82
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v15, v85, 16, v86
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v83, 0xffff, v13
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v86, 0xffff, v11
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff, v96
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v96, 0xffff, v10
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v85, 0xffff, v12
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v12, v87, 16, v83
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v10, v9, 16, v86
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v13, v97, 16, v82
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v82, 0xffff, v7
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v9, v81, 16, v96
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v81, 0xffff, v8
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v83, 0xffff, v6
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v11, v84, 16, v85
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v6, v69, 16, v82
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v69, 0xffff, v2
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v7, v70, 16, v81
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v70, 0xffff, v0
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v2, v64, 16, v66
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v64, 0xffff, v29
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v0, v55, 16, v69
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v55, 0xffff, v30
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v66, 0xffff, v27
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v27, v50, 16, v65
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v28, v51, 16, v64
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v50, 0xffff, v25
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v29, v52, 16, v55
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v51, 0xffff, v24
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v52, 0xffff, v23
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v30, v53, 16, v54
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v53, 0xffff, v22
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v25, v48, 16, v49
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v24, v39, 16, v50
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v23, v38, 16, v51
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v22, v37, 16, v52
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v37, 0xffff, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v38, 0xffff, v19
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v39, 0xffff, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v48, 0xffff, v17
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v8, v71, 16, v80
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v4, v4, 16, v83
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v31, v31, 16, v70
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v26, v26, 16, v66
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v21, v21, 16, v53
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v20, v35, 16, v36
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v19, v34, 16, v37
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v18, v33, 16, v38
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v17, v32, 16, v39
; GFX11-FAKE16-NEXT:    v_lshl_or_b32 v16, v16, 16, v48
; GFX11-FAKE16-NEXT:    s_branch .LBB101_5
; GFX11-FAKE16-NEXT:  .LBB101_3:
; GFX11-FAKE16-NEXT:    s_branch .LBB101_2
; GFX11-FAKE16-NEXT:  .LBB101_4:
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-FAKE16-NEXT:  .LBB101_5: ; %end
; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define <64 x bfloat> @bitcast_v64f16_to_v64bf16(<64 x half> %a, i32 %b) {
; SI-LABEL: bitcast_v64f16_to_v64bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v30
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v22
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v20
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v18
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v16
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v5
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v2
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v1
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v60
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB102_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v58
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v59
; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v60
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v61
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v32
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v34
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v35
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v38
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v48
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v49
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v50
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v51
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v52
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v53
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v17
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v54
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v55
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v19
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v40
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v21
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v42
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v23
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v25
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v56
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v29
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v30
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; kill: killed $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:  .LBB102_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB102_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v61
; SI-NEXT:    v_cvt_f32_f16_e32 v33, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v59
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v58
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_add_f32_e32 v59, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v31
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v29
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v57
; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v57, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v28
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v56
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v27
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v47
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v47, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v26
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v46
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v46, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v25
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v45
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v45, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v24
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v44
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v44, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v43
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_add_f32_e32 v43, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v22
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v42
; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v42, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v21
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v41
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v41, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v20
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v40
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
; SI-NEXT:    v_add_f32_e32 v40, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v19
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v55
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v55, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v54
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v54, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v53
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_add_f32_e32 v53, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v52
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v52, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v51
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v51, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v14
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v50
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_add_f32_e32 v50, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v13
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v49
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v12
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v48
; SI-NEXT:    v_cvt_f32_f16_e32 v34, v34
; SI-NEXT:    v_add_f32_e32 v48, 0x38000000, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v11
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v58
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v58, 0x38000000, v60
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v36, v36
; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
; SI-NEXT:    v_cvt_f32_f16_e32 v37, v37
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v38, v38
; SI-NEXT:    v_add_f32_e32 v36, 0x38000000, v36
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v36, v36
; SI-NEXT:    v_cvt_f32_f16_e32 v39, v39
; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v38, v38
; SI-NEXT:    v_add_f32_e32 v39, 0x38000000, v39
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v39, v39
; SI-NEXT:    v_cvt_f16_f32_e32 v58, v58
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v48, v48
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v50, v50
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v51, v51
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v52, v52
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v53, v53
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v54, v54
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v55, v55
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_cvt_f32_f16_e32 v61, v61
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_cvt_f32_f16_e32 v62, v62
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_cvt_f32_f16_e32 v63, v63
; SI-NEXT:    v_cvt_f16_f32_e32 v40, v40
; SI-NEXT:    v_add_f32_e32 v61, 0x38000000, v61
; SI-NEXT:    v_add_f32_e32 v62, 0x38000000, v62
; SI-NEXT:    v_add_f32_e32 v63, 0x38000000, v63
; SI-NEXT:    v_cvt_f16_f32_e32 v63, v63
; SI-NEXT:    v_cvt_f16_f32_e32 v62, v62
; SI-NEXT:    v_cvt_f16_f32_e32 v61, v61
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v63
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v62
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v61
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v32
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v60
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f32_e32 v60, 0x38000000, v60
; SI-NEXT:    v_cvt_f16_f32_e32 v60, v60
; SI-NEXT:    v_cvt_f16_f32_e32 v41, v41
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v42, v42
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v60
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v34
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v35
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v38
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v48
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v49
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v50
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v51
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v52
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v53
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v17
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v54
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v55
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v19
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v40
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v43, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v41
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v44, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v42
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v45, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v43
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v46, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v44
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v47, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v45
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v25
; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v46
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v57, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v47
; SI-NEXT:    v_cvt_f16_f32_e32 v59, v59
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v56
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v63, 16, v29
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v30
; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v59
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:  .LBB102_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v62
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_alignbit_b32 v18, v18, v19, 16
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_alignbit_b32 v19, v19, v20, 16
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_alignbit_b32 v21, v21, v22, 16
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_alignbit_b32 v22, v22, v23, 16
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_alignbit_b32 v23, v23, v24, 16
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_alignbit_b32 v24, v24, v25, 16
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_alignbit_b32 v25, v25, v26, 16
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_alignbit_b32 v26, v26, v27, 16
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_alignbit_b32 v27, v27, v28, 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_alignbit_b32 v28, v28, v29, 16
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v63
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_alignbit_b32 v29, v29, v30, 16
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_alignbit_b32 v30, v30, v31, 16
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v33
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v64bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB102_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x200
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v15
; VI-NEXT:    v_add_f16_sdwa v15, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v33, v15
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v14
; VI-NEXT:    v_add_f16_sdwa v14, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v33, v14
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v13
; VI-NEXT:    v_add_f16_sdwa v13, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v33, v13
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v12
; VI-NEXT:    v_add_f16_sdwa v12, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v33, v12
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v11
; VI-NEXT:    v_add_f16_sdwa v11, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v33, v11
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v10
; VI-NEXT:    v_add_f16_sdwa v10, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v33, v10
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v9
; VI-NEXT:    v_add_f16_sdwa v9, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v33, v9
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v8
; VI-NEXT:    v_add_f16_sdwa v8, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v33, v8
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v7
; VI-NEXT:    v_add_f16_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v33, v7
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v6
; VI-NEXT:    v_add_f16_sdwa v6, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v33, v6
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v5
; VI-NEXT:    v_add_f16_sdwa v5, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v33, v5
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v4
; VI-NEXT:    v_add_f16_sdwa v4, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v33, v4
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v3
; VI-NEXT:    v_add_f16_sdwa v3, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v33, v3
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v2
; VI-NEXT:    v_add_f16_sdwa v2, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v33, v2
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v1
; VI-NEXT:    v_add_f16_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v33, v1
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v0
; VI-NEXT:    v_add_f16_sdwa v0, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v33, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v31
; VI-NEXT:    v_add_f16_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v33, v31
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v30
; VI-NEXT:    v_add_f16_sdwa v30, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v33, v30
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v29
; VI-NEXT:    v_add_f16_sdwa v29, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v33, v29
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v28
; VI-NEXT:    v_add_f16_sdwa v28, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v33, v28
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v27
; VI-NEXT:    v_add_f16_sdwa v27, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v33, v27
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v26
; VI-NEXT:    v_add_f16_sdwa v26, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v33, v26
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v25
; VI-NEXT:    v_add_f16_sdwa v25, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v33, v25
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v24
; VI-NEXT:    v_add_f16_sdwa v24, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v33, v24
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v23
; VI-NEXT:    v_add_f16_sdwa v23, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v33, v23
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v22
; VI-NEXT:    v_add_f16_sdwa v22, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v33, v22
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v21
; VI-NEXT:    v_add_f16_sdwa v21, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v33, v21
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v20
; VI-NEXT:    v_add_f16_sdwa v20, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v33, v20
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v19
; VI-NEXT:    v_add_f16_sdwa v19, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v33, v19
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v18
; VI-NEXT:    v_add_f16_sdwa v18, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v33, v18
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v17
; VI-NEXT:    v_add_f16_sdwa v17, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, 0x200, v16
; VI-NEXT:    v_or_b32_e32 v17, v33, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB102_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v64bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB102_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_movk_i32 s6, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB102_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v64bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB102_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-NEXT:  .LBB102_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define inreg <64 x bfloat> @bitcast_v64f16_to_v64bf16_scalar(<64 x half> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    v_writelane_b32 v63, s48, 8
; SI-NEXT:    v_writelane_b32 v63, s49, 9
; SI-NEXT:    v_writelane_b32 v63, s50, 10
; SI-NEXT:    v_writelane_b32 v63, s51, 11
; SI-NEXT:    v_writelane_b32 v63, s52, 12
; SI-NEXT:    v_writelane_b32 v63, s53, 13
; SI-NEXT:    v_writelane_b32 v63, s54, 14
; SI-NEXT:    v_writelane_b32 v63, s55, 15
; SI-NEXT:    v_writelane_b32 v63, s64, 16
; SI-NEXT:    v_writelane_b32 v63, s65, 17
; SI-NEXT:    v_writelane_b32 v63, s66, 18
; SI-NEXT:    v_writelane_b32 v63, s67, 19
; SI-NEXT:    v_writelane_b32 v63, s68, 20
; SI-NEXT:    v_writelane_b32 v63, s69, 21
; SI-NEXT:    v_writelane_b32 v63, s70, 22
; SI-NEXT:    v_writelane_b32 v63, s71, 23
; SI-NEXT:    v_writelane_b32 v63, s80, 24
; SI-NEXT:    v_writelane_b32 v63, s81, 25
; SI-NEXT:    v_writelane_b32 v63, s82, 26
; SI-NEXT:    s_lshr_b32 s5, s16, 16
; SI-NEXT:    ; implicit-def: $vgpr62 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v63, s83, 27
; SI-NEXT:    s_lshr_b32 s6, s17, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v62, s5, 0
; SI-NEXT:    v_writelane_b32 v63, s84, 28
; SI-NEXT:    s_lshr_b32 s7, s18, 16
; SI-NEXT:    v_writelane_b32 v62, s6, 1
; SI-NEXT:    v_writelane_b32 v63, s85, 29
; SI-NEXT:    s_lshr_b32 s8, s19, 16
; SI-NEXT:    v_writelane_b32 v62, s7, 2
; SI-NEXT:    v_writelane_b32 v63, s86, 30
; SI-NEXT:    s_lshr_b32 s9, s20, 16
; SI-NEXT:    v_writelane_b32 v62, s8, 3
; SI-NEXT:    v_writelane_b32 v63, s87, 31
; SI-NEXT:    s_lshr_b32 s10, s21, 16
; SI-NEXT:    v_writelane_b32 v62, s9, 4
; SI-NEXT:    v_writelane_b32 v63, s96, 32
; SI-NEXT:    s_lshr_b32 s11, s22, 16
; SI-NEXT:    v_writelane_b32 v62, s10, 5
; SI-NEXT:    v_writelane_b32 v63, s97, 33
; SI-NEXT:    s_lshr_b32 s12, s23, 16
; SI-NEXT:    v_readfirstlane_b32 s52, v17
; SI-NEXT:    v_readfirstlane_b32 s48, v16
; SI-NEXT:    v_readfirstlane_b32 s35, v15
; SI-NEXT:    v_readfirstlane_b32 s30, v14
; SI-NEXT:    v_readfirstlane_b32 s94, v13
; SI-NEXT:    v_readfirstlane_b32 s92, v12
; SI-NEXT:    v_readfirstlane_b32 s90, v11
; SI-NEXT:    v_readfirstlane_b32 s88, v10
; SI-NEXT:    v_readfirstlane_b32 s78, v9
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    v_readfirstlane_b32 s74, v7
; SI-NEXT:    v_readfirstlane_b32 s72, v6
; SI-NEXT:    v_readfirstlane_b32 s62, v5
; SI-NEXT:    v_readfirstlane_b32 s60, v4
; SI-NEXT:    v_readfirstlane_b32 s58, v3
; SI-NEXT:    v_readfirstlane_b32 s56, v2
; SI-NEXT:    v_readfirstlane_b32 s46, v1
; SI-NEXT:    v_readfirstlane_b32 s44, v0
; SI-NEXT:    v_writelane_b32 v62, s11, 6
; SI-NEXT:    v_writelane_b32 v63, s98, 34
; SI-NEXT:    s_lshr_b32 s43, s29, 16
; SI-NEXT:    s_lshr_b32 s42, s28, 16
; SI-NEXT:    s_lshr_b32 s41, s27, 16
; SI-NEXT:    s_lshr_b32 s40, s26, 16
; SI-NEXT:    s_lshr_b32 s15, s25, 16
; SI-NEXT:    s_lshr_b32 s13, s24, 16
; SI-NEXT:    s_lshr_b32 s55, s52, 16
; SI-NEXT:    s_lshr_b32 s51, s48, 16
; SI-NEXT:    s_lshr_b32 s38, s35, 16
; SI-NEXT:    s_lshr_b32 s31, s30, 16
; SI-NEXT:    s_lshr_b32 s95, s94, 16
; SI-NEXT:    s_lshr_b32 s93, s92, 16
; SI-NEXT:    s_lshr_b32 s91, s90, 16
; SI-NEXT:    s_lshr_b32 s89, s88, 16
; SI-NEXT:    s_lshr_b32 s79, s78, 16
; SI-NEXT:    s_lshr_b32 s77, s76, 16
; SI-NEXT:    s_lshr_b32 s75, s74, 16
; SI-NEXT:    s_lshr_b32 s73, s72, 16
; SI-NEXT:    s_lshr_b32 s63, s62, 16
; SI-NEXT:    s_lshr_b32 s61, s60, 16
; SI-NEXT:    s_lshr_b32 s59, s58, 16
; SI-NEXT:    s_lshr_b32 s57, s56, 16
; SI-NEXT:    s_lshr_b32 s47, s46, 16
; SI-NEXT:    s_lshr_b32 s45, s44, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v62, s12, 7
; SI-NEXT:    v_writelane_b32 v63, s99, 35
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v62, s13, 8
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB103_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshl_b32 s4, s16, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 35
; SI-NEXT:    s_lshl_b32 s4, s5, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 34
; SI-NEXT:    s_lshl_b32 s4, s17, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 33
; SI-NEXT:    s_lshl_b32 s4, s6, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 32
; SI-NEXT:    s_lshl_b32 s4, s18, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 31
; SI-NEXT:    s_lshl_b32 s4, s7, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 30
; SI-NEXT:    s_lshl_b32 s4, s19, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 29
; SI-NEXT:    s_lshl_b32 s4, s8, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 28
; SI-NEXT:    s_lshl_b32 s4, s20, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 27
; SI-NEXT:    s_lshl_b32 s4, s9, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 26
; SI-NEXT:    s_lshl_b32 s4, s21, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 25
; SI-NEXT:    s_lshl_b32 s4, s10, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 24
; SI-NEXT:    s_lshl_b32 s4, s22, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 23
; SI-NEXT:    s_lshl_b32 s4, s11, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 22
; SI-NEXT:    s_lshl_b32 s4, s23, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 21
; SI-NEXT:    s_lshl_b32 s4, s12, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 20
; SI-NEXT:    s_lshl_b32 s4, s24, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 19
; SI-NEXT:    s_lshl_b32 s4, s13, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 18
; SI-NEXT:    s_lshl_b32 s4, s25, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 17
; SI-NEXT:    s_lshl_b32 s4, s15, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 16
; SI-NEXT:    s_lshl_b32 s4, s26, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 15
; SI-NEXT:    s_lshl_b32 s4, s40, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 14
; SI-NEXT:    s_lshl_b32 s4, s27, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 13
; SI-NEXT:    s_lshl_b32 s4, s41, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 12
; SI-NEXT:    s_lshl_b32 s4, s28, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 11
; SI-NEXT:    s_lshl_b32 s4, s42, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 10
; SI-NEXT:    s_lshl_b32 s4, s29, 16
; SI-NEXT:    v_writelane_b32 v62, s4, 9
; SI-NEXT:    s_lshl_b32 s99, s43, 16
; SI-NEXT:    s_lshl_b32 s34, s44, 16
; SI-NEXT:    s_lshl_b32 s36, s45, 16
; SI-NEXT:    s_lshl_b32 s37, s46, 16
; SI-NEXT:    s_lshl_b32 s39, s47, 16
; SI-NEXT:    s_lshl_b32 s49, s56, 16
; SI-NEXT:    s_lshl_b32 s50, s57, 16
; SI-NEXT:    s_lshl_b32 s53, s58, 16
; SI-NEXT:    s_lshl_b32 s54, s59, 16
; SI-NEXT:    s_lshl_b32 s64, s60, 16
; SI-NEXT:    s_lshl_b32 s65, s61, 16
; SI-NEXT:    s_lshl_b32 s66, s62, 16
; SI-NEXT:    s_lshl_b32 s67, s63, 16
; SI-NEXT:    s_lshl_b32 s6, s72, 16
; SI-NEXT:    s_lshl_b32 s68, s73, 16
; SI-NEXT:    s_lshl_b32 s7, s74, 16
; SI-NEXT:    s_lshl_b32 s69, s75, 16
; SI-NEXT:    s_lshl_b32 s70, s76, 16
; SI-NEXT:    s_lshl_b32 s71, s77, 16
; SI-NEXT:    s_lshl_b32 s8, s78, 16
; SI-NEXT:    s_lshl_b32 s80, s79, 16
; SI-NEXT:    s_lshl_b32 s81, s88, 16
; SI-NEXT:    s_lshl_b32 s9, s89, 16
; SI-NEXT:    s_lshl_b32 s82, s90, 16
; SI-NEXT:    s_lshl_b32 s83, s91, 16
; SI-NEXT:    s_lshl_b32 s10, s92, 16
; SI-NEXT:    s_lshl_b32 s84, s93, 16
; SI-NEXT:    s_lshl_b32 s85, s94, 16
; SI-NEXT:    s_lshl_b32 s11, s95, 16
; SI-NEXT:    s_lshl_b32 s86, s30, 16
; SI-NEXT:    s_lshl_b32 s87, s31, 16
; SI-NEXT:    s_lshl_b32 s12, s35, 16
; SI-NEXT:    s_lshl_b32 s96, s38, 16
; SI-NEXT:    s_lshl_b32 s13, s48, 16
; SI-NEXT:    s_lshl_b32 s97, s51, 16
; SI-NEXT:    s_lshl_b32 s98, s52, 16
; SI-NEXT:    s_lshl_b32 s14, s55, 16
; SI-NEXT:    s_cbranch_execnz .LBB103_4
; SI-NEXT:  .LBB103_2: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s55
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s52
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s51
; SI-NEXT:    v_readlane_b32 s4, v62, 8
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s48
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s38
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s35
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s31
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s30
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s95
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s94
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s93
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v51, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s91
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s92
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s89
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v48, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s79
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s90
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s77
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s88
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v53, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s75
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v49, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s78
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v55, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s73
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v52, v0
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v55
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s76
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v41, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s63
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v54, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s74
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v41
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v48
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v40, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s72
; SI-NEXT:    v_cvt_f16_f32_e32 v43, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s61
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v42, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v45, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s62
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v52
; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v45
; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v53
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v51
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v44, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s60
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s59
; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v43
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v44
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v46, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s58
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v35, v1
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v46
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v40
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v49
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s57
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v34, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s56
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s47
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s46
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s45
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s44
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s43
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s29
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s42
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s28
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s41
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s27
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s40
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s26
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s15
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s25
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s24
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_readlane_b32 s4, v62, 7
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v36, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s23
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_readlane_b32 s4, v62, 6
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s22
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_readlane_b32 s4, v62, 5
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s21
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_readlane_b32 s4, v62, 4
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s20
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_readlane_b32 s4, v62, 3
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v37, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s19
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_readlane_b32 s4, v62, 2
; SI-NEXT:    v_cvt_f16_f32_e32 v38, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v0
; SI-NEXT:    v_readlane_b32 s4, v62, 1
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 0
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s17
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s4
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v50, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v39, v1
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v50
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v38
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v37
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v36
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v34
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v54
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v8
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_branch .LBB103_5
; SI-NEXT:  .LBB103_3:
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr99
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr37
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr7
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr71
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr81
; SI-NEXT:    ; implicit-def: $sgpr9
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr83
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr85
; SI-NEXT:    ; implicit-def: $sgpr11
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr87
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr13
; SI-NEXT:    ; implicit-def: $sgpr97
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    s_branch .LBB103_2
; SI-NEXT:  .LBB103_4:
; SI-NEXT:    v_readlane_b32 s4, v62, 9
; SI-NEXT:    v_mov_b32_e32 v26, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 10
; SI-NEXT:    v_mov_b32_e32 v25, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 11
; SI-NEXT:    v_mov_b32_e32 v24, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 12
; SI-NEXT:    v_mov_b32_e32 v23, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 13
; SI-NEXT:    v_mov_b32_e32 v22, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 14
; SI-NEXT:    v_mov_b32_e32 v21, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 15
; SI-NEXT:    v_mov_b32_e32 v20, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 16
; SI-NEXT:    v_mov_b32_e32 v19, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 17
; SI-NEXT:    v_mov_b32_e32 v18, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 18
; SI-NEXT:    v_mov_b32_e32 v17, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 19
; SI-NEXT:    v_mov_b32_e32 v50, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 20
; SI-NEXT:    v_mov_b32_e32 v15, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 21
; SI-NEXT:    v_mov_b32_e32 v14, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 22
; SI-NEXT:    v_mov_b32_e32 v13, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 23
; SI-NEXT:    v_mov_b32_e32 v12, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 24
; SI-NEXT:    v_mov_b32_e32 v11, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 25
; SI-NEXT:    v_mov_b32_e32 v10, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 26
; SI-NEXT:    v_mov_b32_e32 v9, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 27
; SI-NEXT:    v_mov_b32_e32 v38, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 28
; SI-NEXT:    v_mov_b32_e32 v0, s14
; SI-NEXT:    v_mov_b32_e32 v39, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 29
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s98
; SI-NEXT:    v_mov_b32_e32 v6, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 30
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s97
; SI-NEXT:    v_mov_b32_e32 v5, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 31
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s13
; SI-NEXT:    v_mov_b32_e32 v4, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 32
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s96
; SI-NEXT:    v_mov_b32_e32 v3, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 33
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s12
; SI-NEXT:    v_mov_b32_e32 v2, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 34
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s86
; SI-NEXT:    v_mov_b32_e32 v1, s4
; SI-NEXT:    v_readlane_b32 s4, v62, 35
; SI-NEXT:    v_mov_b32_e32 v49, s87
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v51, s11
; SI-NEXT:    v_mov_b32_e32 v48, s85
; SI-NEXT:    v_mov_b32_e32 v53, s84
; SI-NEXT:    v_mov_b32_e32 v52, s10
; SI-NEXT:    v_mov_b32_e32 v55, s83
; SI-NEXT:    v_mov_b32_e32 v54, s82
; SI-NEXT:    v_mov_b32_e32 v41, s9
; SI-NEXT:    v_mov_b32_e32 v40, s81
; SI-NEXT:    v_mov_b32_e32 v43, s80
; SI-NEXT:    v_mov_b32_e32 v42, s8
; SI-NEXT:    v_mov_b32_e32 v45, s71
; SI-NEXT:    v_mov_b32_e32 v44, s70
; SI-NEXT:    v_mov_b32_e32 v47, s69
; SI-NEXT:    v_mov_b32_e32 v46, s7
; SI-NEXT:    v_mov_b32_e32 v57, s68
; SI-NEXT:    v_mov_b32_e32 v56, s6
; SI-NEXT:    v_mov_b32_e32 v59, s67
; SI-NEXT:    v_mov_b32_e32 v58, s66
; SI-NEXT:    v_mov_b32_e32 v61, s65
; SI-NEXT:    v_mov_b32_e32 v60, s64
; SI-NEXT:    v_mov_b32_e32 v37, s54
; SI-NEXT:    v_mov_b32_e32 v36, s53
; SI-NEXT:    v_mov_b32_e32 v33, s50
; SI-NEXT:    v_mov_b32_e32 v32, s49
; SI-NEXT:    v_mov_b32_e32 v31, s39
; SI-NEXT:    v_mov_b32_e32 v30, s37
; SI-NEXT:    v_mov_b32_e32 v29, s36
; SI-NEXT:    v_mov_b32_e32 v28, s34
; SI-NEXT:    v_mov_b32_e32 v27, s99
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s4
; SI-NEXT:  .LBB103_5: ; %end
; SI-NEXT:    v_mul_f32_e32 v34, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v5
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v39
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v9
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v11
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v13
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v17
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v19
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v21
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v14
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v23
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v25
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v27
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v29
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v18
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v37
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v20
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v50, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v4
; SI-NEXT:    v_lshr_b64 v[33:34], v[34:35], 16
; SI-NEXT:    v_lshr_b64 v[34:35], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[2:3], v[7:8], 16
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v38
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v36
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v60
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v58
; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v46
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v44
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v42
; SI-NEXT:    v_mul_f32_e32 v40, 1.0, v40
; SI-NEXT:    v_mul_f32_e32 v54, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v52, 1.0, v52
; SI-NEXT:    v_readlane_b32 s99, v63, 35
; SI-NEXT:    v_readlane_b32 s98, v63, 34
; SI-NEXT:    v_readlane_b32 s97, v63, 33
; SI-NEXT:    v_readlane_b32 s96, v63, 32
; SI-NEXT:    v_readlane_b32 s87, v63, 31
; SI-NEXT:    v_readlane_b32 s86, v63, 30
; SI-NEXT:    v_readlane_b32 s85, v63, 29
; SI-NEXT:    v_readlane_b32 s84, v63, 28
; SI-NEXT:    v_readlane_b32 s83, v63, 27
; SI-NEXT:    v_readlane_b32 s82, v63, 26
; SI-NEXT:    v_readlane_b32 s81, v63, 25
; SI-NEXT:    v_readlane_b32 s80, v63, 24
; SI-NEXT:    v_readlane_b32 s71, v63, 23
; SI-NEXT:    v_readlane_b32 s70, v63, 22
; SI-NEXT:    v_readlane_b32 s69, v63, 21
; SI-NEXT:    v_readlane_b32 s68, v63, 20
; SI-NEXT:    v_readlane_b32 s67, v63, 19
; SI-NEXT:    v_readlane_b32 s66, v63, 18
; SI-NEXT:    v_readlane_b32 s65, v63, 17
; SI-NEXT:    v_readlane_b32 s64, v63, 16
; SI-NEXT:    v_readlane_b32 s55, v63, 15
; SI-NEXT:    v_readlane_b32 s54, v63, 14
; SI-NEXT:    v_readlane_b32 s53, v63, 13
; SI-NEXT:    v_readlane_b32 s52, v63, 12
; SI-NEXT:    v_readlane_b32 s51, v63, 11
; SI-NEXT:    v_readlane_b32 s50, v63, 10
; SI-NEXT:    v_readlane_b32 s49, v63, 9
; SI-NEXT:    v_readlane_b32 s48, v63, 8
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v48, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    v_lshr_b64 v[4:5], v[5:6], 16
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    v_lshr_b64 v[7:8], v[15:16], 16
; SI-NEXT:    v_lshr_b64 v[8:9], v[9:10], 16
; SI-NEXT:    v_lshr_b64 v[9:10], v[11:12], 16
; SI-NEXT:    v_lshr_b64 v[10:11], v[13:14], 16
; SI-NEXT:    v_lshr_b64 v[11:12], v[22:23], 16
; SI-NEXT:    v_lshr_b64 v[12:13], v[24:25], 16
; SI-NEXT:    v_lshr_b64 v[13:14], v[26:27], 16
; SI-NEXT:    v_lshr_b64 v[14:15], v[28:29], 16
; SI-NEXT:    v_lshr_b64 v[15:16], v[30:31], 16
; SI-NEXT:    v_lshr_b64 v[16:17], v[17:18], 16
; SI-NEXT:    v_lshr_b64 v[17:18], v[19:20], 16
; SI-NEXT:    v_lshr_b64 v[18:19], v[60:61], 16
; SI-NEXT:    v_lshr_b64 v[19:20], v[58:59], 16
; SI-NEXT:    v_lshr_b64 v[20:21], v[56:57], 16
; SI-NEXT:    v_lshr_b64 v[21:22], v[46:47], 16
; SI-NEXT:    v_lshr_b64 v[22:23], v[44:45], 16
; SI-NEXT:    v_lshr_b64 v[23:24], v[42:43], 16
; SI-NEXT:    v_lshr_b64 v[24:25], v[40:41], 16
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[25:26], v[54:55], 16
; SI-NEXT:    v_lshr_b64 v[26:27], v[52:53], 16
; SI-NEXT:    v_lshr_b64 v[27:28], v[50:51], 16
; SI-NEXT:    v_lshr_b64 v[28:29], v[48:49], 16
; SI-NEXT:    v_lshr_b64 v[29:30], v[38:39], 16
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[30:31], v[36:37], 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_lshr_b64 v[31:32], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v0, v33
; SI-NEXT:    v_mov_b32_e32 v1, v34
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v33, s30, 0
; VI-NEXT:    v_writelane_b32 v33, s31, 1
; VI-NEXT:    v_writelane_b32 v33, s36, 2
; VI-NEXT:    v_writelane_b32 v33, s37, 3
; VI-NEXT:    v_writelane_b32 v33, s38, 4
; VI-NEXT:    v_writelane_b32 v33, s39, 5
; VI-NEXT:    v_writelane_b32 v33, s48, 6
; VI-NEXT:    v_writelane_b32 v33, s49, 7
; VI-NEXT:    v_writelane_b32 v33, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v33, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB103_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB103_4
; VI-NEXT:  .LBB103_2: ; %cmp.true
; VI-NEXT:    s_lshr_b32 s4, s50, 16
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    s_lshr_b32 s4, s30, 16
; VI-NEXT:    v_mov_b32_e32 v16, 0x200
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    s_lshr_b32 s4, s31, 16
; VI-NEXT:    v_add_f16_sdwa v18, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, s30, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_mov_b32_e32 v3, s4
; VI-NEXT:    s_lshr_b32 s4, s29, 16
; VI-NEXT:    v_or_b32_e32 v14, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s29, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s28, 16
; VI-NEXT:    v_or_b32_e32 v13, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s28, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s27, 16
; VI-NEXT:    v_or_b32_e32 v12, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s27, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s26, 16
; VI-NEXT:    v_or_b32_e32 v11, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s26, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s25, 16
; VI-NEXT:    v_or_b32_e32 v10, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s25, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s24, 16
; VI-NEXT:    v_or_b32_e32 v9, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s24, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s23, 16
; VI-NEXT:    v_or_b32_e32 v8, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s23, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s22, 16
; VI-NEXT:    v_or_b32_e32 v7, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s22, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s21, 16
; VI-NEXT:    v_or_b32_e32 v6, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s21, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s20, 16
; VI-NEXT:    v_or_b32_e32 v5, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s20, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s19, 16
; VI-NEXT:    v_or_b32_e32 v4, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v2, s31, v16
; VI-NEXT:    v_add_f16_sdwa v3, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, s19, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s18, 16
; VI-NEXT:    v_or_b32_e32 v15, v2, v3
; VI-NEXT:    v_or_b32_e32 v3, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s18, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s17, 16
; VI-NEXT:    v_or_b32_e32 v2, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    s_lshr_b32 s4, s16, 16
; VI-NEXT:    v_mov_b32_e32 v19, s4
; VI-NEXT:    s_lshr_b32 s4, s51, 16
; VI-NEXT:    v_add_f16_e32 v17, s50, v16
; VI-NEXT:    v_mov_b32_e32 v20, s4
; VI-NEXT:    s_lshr_b32 s4, s49, 16
; VI-NEXT:    v_or_b32_e32 v30, v17, v18
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s49, v16
; VI-NEXT:    s_lshr_b32 s4, s48, 16
; VI-NEXT:    v_or_b32_e32 v29, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s48, v16
; VI-NEXT:    s_lshr_b32 s4, s47, 16
; VI-NEXT:    v_or_b32_e32 v28, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s47, v16
; VI-NEXT:    s_lshr_b32 s4, s46, 16
; VI-NEXT:    v_or_b32_e32 v27, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s46, v16
; VI-NEXT:    s_lshr_b32 s4, s45, 16
; VI-NEXT:    v_or_b32_e32 v26, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s45, v16
; VI-NEXT:    s_lshr_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v25, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s44, v16
; VI-NEXT:    s_lshr_b32 s4, s43, 16
; VI-NEXT:    v_or_b32_e32 v24, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s43, v16
; VI-NEXT:    s_lshr_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v23, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s42, v16
; VI-NEXT:    s_lshr_b32 s4, s41, 16
; VI-NEXT:    v_or_b32_e32 v22, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_e32 v0, s17, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s41, v16
; VI-NEXT:    s_lshr_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v1, v0, v1
; VI-NEXT:    v_add_f16_e32 v0, s16, v16
; VI-NEXT:    v_add_f16_sdwa v19, v19, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_or_b32_e32 v0, v0, v19
; VI-NEXT:    v_add_f16_e32 v19, s51, v16
; VI-NEXT:    v_add_f16_sdwa v20, v20, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s40, v16
; VI-NEXT:    s_lshr_b32 s4, s39, 16
; VI-NEXT:    v_or_b32_e32 v31, v19, v20
; VI-NEXT:    v_or_b32_e32 v20, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s39, v16
; VI-NEXT:    s_lshr_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v19, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s38, v16
; VI-NEXT:    s_lshr_b32 s4, s37, 16
; VI-NEXT:    v_or_b32_e32 v18, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v32, s37, v16
; VI-NEXT:    s_lshr_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_mov_b32_e32 v32, s4
; VI-NEXT:    v_add_f16_sdwa v32, v32, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, s36, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:    s_branch .LBB103_5
; VI-NEXT:  .LBB103_3:
; VI-NEXT:    s_branch .LBB103_2
; VI-NEXT:  .LBB103_4:
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:  .LBB103_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v33, 9
; VI-NEXT:    v_readlane_b32 s50, v33, 8
; VI-NEXT:    v_readlane_b32 s49, v33, 7
; VI-NEXT:    v_readlane_b32 s48, v33, 6
; VI-NEXT:    v_readlane_b32 s39, v33, 5
; VI-NEXT:    v_readlane_b32 s38, v33, 4
; VI-NEXT:    v_readlane_b32 s37, v33, 3
; VI-NEXT:    v_readlane_b32 s36, v33, 2
; VI-NEXT:    v_readlane_b32 s31, v33, 1
; VI-NEXT:    v_readlane_b32 s30, v33, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB103_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB103_4
; GFX9-NEXT:  .LBB103_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, s31, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, s30, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, s29, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, s28, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, s27, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, s26, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, s25, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, s24, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, s23, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, s22, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, s21, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, s20, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, s19, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, s18, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, s17, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, s16, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v31, s51, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, s50, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, s49, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, s48, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, s47, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, s46, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, s45, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, s44, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, s43, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, s42, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, s41, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, s40, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, s39, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, s38, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, s37, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, s36, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB103_5
; GFX9-NEXT:  .LBB103_3:
; GFX9-NEXT:    s_branch .LBB103_2
; GFX9-NEXT:  .LBB103_4:
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:  .LBB103_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v64bf16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB103_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB103_4
; GFX11-NEXT:  .LBB103_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s16 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, s51 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, s50 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s49 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s48 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, s47 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, s46 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, s45 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s44 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s43 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, s42 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s41 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s40 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s39 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s38 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s37 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s36 op_sel_hi:[0,1]
; GFX11-NEXT:    s_branch .LBB103_5
; GFX11-NEXT:  .LBB103_3:
; GFX11-NEXT:    s_branch .LBB103_2
; GFX11-NEXT:  .LBB103_4:
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:  .LBB103_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define <64 x i16> @bitcast_v64bf16_to_v64i16(<64 x bfloat> %a, i32 %b) {
; SI-LABEL: bitcast_v64bf16_to_v64i16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v19
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v21
; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v11
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v21
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v13
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v29
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v15
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v5
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v36, 0xffff0000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v25
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v17
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v30
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v30
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v28
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v28
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v27
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v27
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v26
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v23
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v23
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    v_and_b32_e32 v58, 0xffff0000, v20
; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v20
; SI-NEXT:    v_and_b32_e32 v63, 0xffff0000, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v12
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v32
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v6
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v8
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v11
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v34
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v13
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v18
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v4
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v5
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v41
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v7
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v36
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v29
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v9
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v17
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v25
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v62
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v0
; SI-NEXT:    v_mul_f32_e32 v38, 1.0, v38
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v37
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v40
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v35
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v10
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v20
; SI-NEXT:    v_mul_f32_e32 v37, 1.0, v12
; SI-NEXT:    v_mul_f32_e32 v41, 1.0, v27
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v14
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v33
; SI-NEXT:    v_mul_f32_e32 v36, 1.0, v16
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v63
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v61
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v60
; SI-NEXT:    v_mul_f32_e32 v43, 1.0, v43
; SI-NEXT:    v_mul_f32_e32 v42, 1.0, v42
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v59
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v58
; SI-NEXT:    v_mul_f32_e32 v45, 1.0, v45
; SI-NEXT:    v_mul_f32_e32 v44, 1.0, v44
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v22
; SI-NEXT:    v_mul_f32_e32 v35, 1.0, v47
; SI-NEXT:    v_mul_f32_e32 v47, 1.0, v46
; SI-NEXT:    v_mul_f32_e32 v46, 1.0, v30
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v24
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_mul_f32_e32 v57, 1.0, v57
; SI-NEXT:    v_mul_f32_e32 v56, 1.0, v56
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v26
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v55
; SI-NEXT:    v_mul_f32_e32 v59, 1.0, v54
; SI-NEXT:    v_mul_f32_e32 v58, 1.0, v53
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v52
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v51
; SI-NEXT:    v_mul_f32_e32 v61, 1.0, v50
; SI-NEXT:    v_mul_f32_e32 v60, 1.0, v49
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v48
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v39
; SI-NEXT:    v_mul_f32_e32 v63, 1.0, v31
; SI-NEXT:    v_mul_f32_e32 v62, 1.0, v23
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr52
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; implicit-def: $vgpr51
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; kill: killed $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; kill: killed $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB104_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v38
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; kill: killed $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v21
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; kill: killed $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; kill: killed $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v32
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; kill: killed $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v37
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v29
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v42
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v45
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v35
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v47
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v30
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v57
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v56
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v26
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v24
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v61
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v60
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v22
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v18
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v63
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v62
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr58
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr60
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:  .LBB104_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB104_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v38
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v15
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v19
; SI-NEXT:    v_alignbit_b32 v0, v15, v0, 16
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; SI-NEXT:    v_alignbit_b32 v0, v2, v0, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v21
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v2
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
; SI-NEXT:    v_alignbit_b32 v0, v2, v0, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v32
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v2
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v5
; SI-NEXT:    v_alignbit_b32 v0, v2, v0, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v2
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v7
; SI-NEXT:    v_alignbit_b32 v0, v2, v0, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v2
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v9
; SI-NEXT:    v_alignbit_b32 v0, v2, v0, 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v37
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
; SI-NEXT:    v_alignbit_b32 v0, v4, v0, 16
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v11
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v4
; SI-NEXT:    v_alignbit_b32 v0, v6, v0, 16
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v29
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v36
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v6
; SI-NEXT:    v_alignbit_b32 v0, v11, v0, 16
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v11
; SI-NEXT:    v_alignbit_b32 v0, v13, v0, 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v10
; SI-NEXT:    v_alignbit_b32 v0, v12, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v8
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v35
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v8
; SI-NEXT:    v_alignbit_b32 v0, v12, v0, 16
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; SI-NEXT:    v_alignbit_b32 v0, v13, v0, 16
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; SI-NEXT:    v_alignbit_b32 v0, v14, v0, 16
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v26
; SI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; SI-NEXT:    v_alignbit_b32 v0, v15, v0, 16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v22
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; SI-NEXT:    v_alignbit_b32 v0, v16, v0, 16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v62
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v63
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v18, v16, v0, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v60
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v61
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v20, v16, v15, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v58
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v59
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v21, v16, v15, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v56
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v57
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v22, v16, v15, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v46
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v47
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v24, v16, v15, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v44
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v45
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v26, v16, v15, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v42
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    v_alignbit_b32 v27, v16, v15, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; SI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; SI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; SI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; SI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; SI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; SI-NEXT:    v_alignbit_b32 v0, v18, v0, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v33, v50, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v34, v51, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v25, v52, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v28, v39, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v23, v53, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    v_alignbit_b32 v31, v48, v15, 16
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v40, v54, v16, 16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v40, v1, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v23, v3, 16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v25, v5, 16
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v34, v7, 16
; SI-NEXT:    v_alignbit_b32 v15, v49, v15, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v33, v9, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v15, v2, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v31, v4, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; SI-NEXT:    v_alignbit_b32 v17, v55, v16, 16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v19
; SI-NEXT:    v_mov_b32_e32 v19, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v19, v6, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v27, v11, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v26, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v26, v10, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v24, v8, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v22, v12, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v21, v13, 16
; SI-NEXT:    v_alignbit_b32 v16, v17, v16, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v20, v14, 16
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:  .LBB104_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v55
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v54
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v53
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v52
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v51
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v50
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v49
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v48
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v39
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v40
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v34
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v33
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v64i16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; VI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; VI-NEXT:    s_waitcnt vmcnt(14)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB104_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; VI-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; VI-NEXT:    v_bfe_u32 v33, v32, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v32
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; VI-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; VI-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; VI-NEXT:    v_bfe_u32 v33, v16, 16, 1
; VI-NEXT:    s_movk_i32 s6, 0x7fff
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v16
; VI-NEXT:    v_add_u32_e32 v33, vcc, s6, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; VI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; VI-NEXT:    v_bfe_u32 v34, v33, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v33
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; VI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; VI-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; VI-NEXT:    v_bfe_u32 v34, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v34, vcc, v34, v17
; VI-NEXT:    v_add_u32_e32 v34, vcc, s6, v34
; VI-NEXT:    v_or_b32_e32 v35, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; VI-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v34, 16, v18
; VI-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; VI-NEXT:    v_bfe_u32 v35, v34, 16, 1
; VI-NEXT:    v_add_u32_e32 v35, vcc, v35, v34
; VI-NEXT:    v_add_u32_e32 v35, vcc, s6, v35
; VI-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; VI-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; VI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; VI-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc
; VI-NEXT:    v_bfe_u32 v35, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v35, vcc, v35, v18
; VI-NEXT:    v_add_u32_e32 v35, vcc, s6, v35
; VI-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v35, v36, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
; VI-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; VI-NEXT:    v_bfe_u32 v36, v35, 16, 1
; VI-NEXT:    v_add_u32_e32 v36, vcc, v36, v35
; VI-NEXT:    v_add_u32_e32 v36, vcc, s6, v36
; VI-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; VI-NEXT:    v_or_b32_e32 v37, 0x400000, v35
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; VI-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; VI-NEXT:    v_cndmask_b32_e32 v35, v36, v37, vcc
; VI-NEXT:    v_bfe_u32 v36, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v36, vcc, v36, v19
; VI-NEXT:    v_add_u32_e32 v36, vcc, s6, v36
; VI-NEXT:    v_or_b32_e32 v37, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_cndmask_b32_e32 v19, v36, v37, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
; VI-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
; VI-NEXT:    v_bfe_u32 v37, v36, 16, 1
; VI-NEXT:    v_add_u32_e32 v37, vcc, v37, v36
; VI-NEXT:    v_add_u32_e32 v37, vcc, s6, v37
; VI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; VI-NEXT:    v_or_b32_e32 v38, 0x400000, v36
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; VI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; VI-NEXT:    v_cndmask_b32_e32 v36, v37, v38, vcc
; VI-NEXT:    v_bfe_u32 v37, v20, 16, 1
; VI-NEXT:    v_add_u32_e32 v37, vcc, v37, v20
; VI-NEXT:    v_add_u32_e32 v37, vcc, s6, v37
; VI-NEXT:    v_or_b32_e32 v38, 0x400000, v20
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; VI-NEXT:    v_cndmask_b32_e32 v20, v37, v38, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; VI-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
; VI-NEXT:    v_bfe_u32 v38, v37, 16, 1
; VI-NEXT:    v_add_u32_e32 v38, vcc, v38, v37
; VI-NEXT:    v_add_u32_e32 v38, vcc, s6, v38
; VI-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; VI-NEXT:    v_or_b32_e32 v39, 0x400000, v37
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; VI-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; VI-NEXT:    v_cndmask_b32_e32 v37, v38, v39, vcc
; VI-NEXT:    v_bfe_u32 v38, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v38, vcc, v38, v21
; VI-NEXT:    v_add_u32_e32 v38, vcc, s6, v38
; VI-NEXT:    v_or_b32_e32 v39, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_cndmask_b32_e32 v21, v38, v39, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v38, 16, v22
; VI-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; VI-NEXT:    v_bfe_u32 v39, v38, 16, 1
; VI-NEXT:    v_add_u32_e32 v39, vcc, v39, v38
; VI-NEXT:    v_add_u32_e32 v39, vcc, s6, v39
; VI-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; VI-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; VI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; VI-NEXT:    v_cndmask_b32_e32 v38, v39, v48, vcc
; VI-NEXT:    v_bfe_u32 v39, v22, 16, 1
; VI-NEXT:    v_add_u32_e32 v39, vcc, v39, v22
; VI-NEXT:    v_add_u32_e32 v39, vcc, s6, v39
; VI-NEXT:    v_or_b32_e32 v48, 0x400000, v22
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; VI-NEXT:    v_cndmask_b32_e32 v22, v39, v48, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
; VI-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
; VI-NEXT:    v_bfe_u32 v48, v39, 16, 1
; VI-NEXT:    v_add_u32_e32 v48, vcc, v48, v39
; VI-NEXT:    v_add_u32_e32 v48, vcc, s6, v48
; VI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; VI-NEXT:    v_or_b32_e32 v49, 0x400000, v39
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; VI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; VI-NEXT:    v_cndmask_b32_e32 v39, v48, v49, vcc
; VI-NEXT:    v_bfe_u32 v48, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v48, vcc, v48, v23
; VI-NEXT:    v_add_u32_e32 v48, vcc, s6, v48
; VI-NEXT:    v_or_b32_e32 v49, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_cndmask_b32_e32 v23, v48, v49, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; VI-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
; VI-NEXT:    v_bfe_u32 v49, v48, 16, 1
; VI-NEXT:    v_add_u32_e32 v49, vcc, v49, v48
; VI-NEXT:    v_add_u32_e32 v49, vcc, s6, v49
; VI-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; VI-NEXT:    v_or_b32_e32 v50, 0x400000, v48
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; VI-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; VI-NEXT:    v_cndmask_b32_e32 v48, v49, v50, vcc
; VI-NEXT:    v_bfe_u32 v49, v24, 16, 1
; VI-NEXT:    v_add_u32_e32 v49, vcc, v49, v24
; VI-NEXT:    v_add_u32_e32 v49, vcc, s6, v49
; VI-NEXT:    v_or_b32_e32 v50, 0x400000, v24
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; VI-NEXT:    v_cndmask_b32_e32 v24, v49, v50, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v49, 16, v25
; VI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; VI-NEXT:    v_bfe_u32 v50, v49, 16, 1
; VI-NEXT:    v_add_u32_e32 v50, vcc, v50, v49
; VI-NEXT:    v_add_u32_e32 v50, vcc, s6, v50
; VI-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; VI-NEXT:    v_or_b32_e32 v51, 0x400000, v49
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; VI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; VI-NEXT:    v_cndmask_b32_e32 v49, v50, v51, vcc
; VI-NEXT:    v_bfe_u32 v50, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v50, vcc, v50, v25
; VI-NEXT:    v_add_u32_e32 v50, vcc, s6, v50
; VI-NEXT:    v_or_b32_e32 v51, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_cndmask_b32_e32 v25, v50, v51, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v50, 16, v26
; VI-NEXT:    v_add_f32_e32 v50, 0x40c00000, v50
; VI-NEXT:    v_bfe_u32 v51, v50, 16, 1
; VI-NEXT:    v_add_u32_e32 v51, vcc, v51, v50
; VI-NEXT:    v_add_u32_e32 v51, vcc, s6, v51
; VI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; VI-NEXT:    v_or_b32_e32 v52, 0x400000, v50
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; VI-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; VI-NEXT:    v_cndmask_b32_e32 v50, v51, v52, vcc
; VI-NEXT:    v_bfe_u32 v51, v26, 16, 1
; VI-NEXT:    v_add_u32_e32 v51, vcc, v51, v26
; VI-NEXT:    v_add_u32_e32 v51, vcc, s6, v51
; VI-NEXT:    v_or_b32_e32 v52, 0x400000, v26
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; VI-NEXT:    v_cndmask_b32_e32 v26, v51, v52, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v51, 16, v27
; VI-NEXT:    v_add_f32_e32 v51, 0x40c00000, v51
; VI-NEXT:    v_bfe_u32 v52, v51, 16, 1
; VI-NEXT:    v_add_u32_e32 v52, vcc, v52, v51
; VI-NEXT:    v_add_u32_e32 v52, vcc, s6, v52
; VI-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; VI-NEXT:    v_or_b32_e32 v53, 0x400000, v51
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; VI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; VI-NEXT:    v_cndmask_b32_e32 v51, v52, v53, vcc
; VI-NEXT:    v_bfe_u32 v52, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v52, vcc, v52, v27
; VI-NEXT:    v_add_u32_e32 v52, vcc, s6, v52
; VI-NEXT:    v_or_b32_e32 v53, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_cndmask_b32_e32 v27, v52, v53, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v52, 16, v28
; VI-NEXT:    v_add_f32_e32 v52, 0x40c00000, v52
; VI-NEXT:    v_bfe_u32 v53, v52, 16, 1
; VI-NEXT:    v_add_u32_e32 v53, vcc, v53, v52
; VI-NEXT:    v_add_u32_e32 v53, vcc, s6, v53
; VI-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; VI-NEXT:    v_or_b32_e32 v54, 0x400000, v52
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; VI-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; VI-NEXT:    v_cndmask_b32_e32 v52, v53, v54, vcc
; VI-NEXT:    v_bfe_u32 v53, v28, 16, 1
; VI-NEXT:    v_add_u32_e32 v53, vcc, v53, v28
; VI-NEXT:    v_add_u32_e32 v53, vcc, s6, v53
; VI-NEXT:    v_or_b32_e32 v54, 0x400000, v28
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; VI-NEXT:    v_cndmask_b32_e32 v28, v53, v54, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v53, 16, v29
; VI-NEXT:    v_add_f32_e32 v53, 0x40c00000, v53
; VI-NEXT:    v_bfe_u32 v54, v53, 16, 1
; VI-NEXT:    v_add_u32_e32 v54, vcc, v54, v53
; VI-NEXT:    v_add_u32_e32 v54, vcc, s6, v54
; VI-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; VI-NEXT:    v_or_b32_e32 v55, 0x400000, v53
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; VI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; VI-NEXT:    v_cndmask_b32_e32 v53, v54, v55, vcc
; VI-NEXT:    v_bfe_u32 v54, v29, 16, 1
; VI-NEXT:    v_add_u32_e32 v54, vcc, v54, v29
; VI-NEXT:    v_add_u32_e32 v54, vcc, s6, v54
; VI-NEXT:    v_or_b32_e32 v55, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_cndmask_b32_e32 v29, v54, v55, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v54, 16, v30
; VI-NEXT:    v_add_f32_e32 v54, 0x40c00000, v54
; VI-NEXT:    v_bfe_u32 v55, v54, 16, 1
; VI-NEXT:    v_add_u32_e32 v55, vcc, v55, v54
; VI-NEXT:    v_add_u32_e32 v55, vcc, s6, v55
; VI-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; VI-NEXT:    v_or_b32_e32 v40, 0x400000, v54
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; VI-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; VI-NEXT:    v_cndmask_b32_e32 v54, v55, v40, vcc
; VI-NEXT:    v_bfe_u32 v55, v30, 16, 1
; VI-NEXT:    v_add_u32_e32 v55, vcc, v55, v30
; VI-NEXT:    v_add_u32_e32 v55, vcc, s6, v55
; VI-NEXT:    v_or_b32_e32 v40, 0x400000, v30
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; VI-NEXT:    v_cndmask_b32_e32 v30, v55, v40, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v55, 16, v31
; VI-NEXT:    v_add_f32_e32 v55, 0x40c00000, v55
; VI-NEXT:    v_bfe_u32 v40, v55, 16, 1
; VI-NEXT:    v_add_u32_e32 v40, vcc, v40, v55
; VI-NEXT:    v_add_u32_e32 v40, vcc, s6, v40
; VI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; VI-NEXT:    v_or_b32_e32 v41, 0x400000, v55
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; VI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; VI-NEXT:    v_cndmask_b32_e32 v55, v40, v41, vcc
; VI-NEXT:    v_bfe_u32 v40, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v40, vcc, v40, v31
; VI-NEXT:    v_add_u32_e32 v40, vcc, s6, v40
; VI-NEXT:    v_or_b32_e32 v41, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v40, v41, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v40, 16, v0
; VI-NEXT:    v_add_f32_e32 v40, 0x40c00000, v40
; VI-NEXT:    v_bfe_u32 v41, v40, 16, 1
; VI-NEXT:    v_add_u32_e32 v41, vcc, v41, v40
; VI-NEXT:    v_add_u32_e32 v41, vcc, s6, v41
; VI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; VI-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; VI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; VI-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; VI-NEXT:    v_bfe_u32 v41, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v41, vcc, v41, v0
; VI-NEXT:    v_add_u32_e32 v41, vcc, s6, v41
; VI-NEXT:    v_or_b32_e32 v42, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v17, v41, v42, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v41, 16, v1
; VI-NEXT:    v_add_f32_e32 v41, 0x40c00000, v41
; VI-NEXT:    v_bfe_u32 v42, v41, 16, 1
; VI-NEXT:    v_add_u32_e32 v42, vcc, v42, v41
; VI-NEXT:    v_add_u32_e32 v42, vcc, s6, v42
; VI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; VI-NEXT:    v_or_b32_e32 v43, 0x400000, v41
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; VI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; VI-NEXT:    v_cndmask_b32_e32 v41, v42, v43, vcc
; VI-NEXT:    v_bfe_u32 v42, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v42, vcc, v42, v1
; VI-NEXT:    v_add_u32_e32 v42, vcc, s6, v42
; VI-NEXT:    v_or_b32_e32 v43, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; VI-NEXT:    v_cndmask_b32_e32 v16, v42, v43, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v42, 16, v2
; VI-NEXT:    v_add_f32_e32 v42, 0x40c00000, v42
; VI-NEXT:    v_bfe_u32 v43, v42, 16, 1
; VI-NEXT:    v_add_u32_e32 v43, vcc, v43, v42
; VI-NEXT:    v_add_u32_e32 v43, vcc, s6, v43
; VI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; VI-NEXT:    v_or_b32_e32 v44, 0x400000, v42
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v42, v42
; VI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; VI-NEXT:    v_cndmask_b32_e32 v42, v43, v44, vcc
; VI-NEXT:    v_bfe_u32 v43, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v43, vcc, v43, v2
; VI-NEXT:    v_add_u32_e32 v43, vcc, s6, v43
; VI-NEXT:    v_or_b32_e32 v44, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    v_cndmask_b32_e32 v2, v43, v44, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v43, 16, v3
; VI-NEXT:    v_add_f32_e32 v43, 0x40c00000, v43
; VI-NEXT:    v_bfe_u32 v44, v43, 16, 1
; VI-NEXT:    v_add_u32_e32 v44, vcc, v44, v43
; VI-NEXT:    v_add_u32_e32 v44, vcc, s6, v44
; VI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; VI-NEXT:    v_or_b32_e32 v45, 0x400000, v43
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v43, v43
; VI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; VI-NEXT:    v_cndmask_b32_e32 v43, v44, v45, vcc
; VI-NEXT:    v_bfe_u32 v44, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v44, vcc, v44, v3
; VI-NEXT:    v_add_u32_e32 v44, vcc, s6, v44
; VI-NEXT:    v_or_b32_e32 v45, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v44, v45, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; VI-NEXT:    v_add_f32_e32 v44, 0x40c00000, v44
; VI-NEXT:    v_bfe_u32 v45, v44, 16, 1
; VI-NEXT:    v_add_u32_e32 v45, vcc, v45, v44
; VI-NEXT:    v_add_u32_e32 v45, vcc, s6, v45
; VI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; VI-NEXT:    v_or_b32_e32 v46, 0x400000, v44
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v44, v44
; VI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; VI-NEXT:    v_cndmask_b32_e32 v44, v45, v46, vcc
; VI-NEXT:    v_bfe_u32 v45, v4, 16, 1
; VI-NEXT:    v_add_u32_e32 v45, vcc, v45, v4
; VI-NEXT:    v_add_u32_e32 v45, vcc, s6, v45
; VI-NEXT:    v_or_b32_e32 v46, 0x400000, v4
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; VI-NEXT:    v_cndmask_b32_e32 v4, v45, v46, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v45, 16, v5
; VI-NEXT:    v_add_f32_e32 v45, 0x40c00000, v45
; VI-NEXT:    v_bfe_u32 v46, v45, 16, 1
; VI-NEXT:    v_add_u32_e32 v46, vcc, v46, v45
; VI-NEXT:    v_add_u32_e32 v46, vcc, s6, v46
; VI-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; VI-NEXT:    v_or_b32_e32 v47, 0x400000, v45
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v45, v45
; VI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; VI-NEXT:    v_cndmask_b32_e32 v45, v46, v47, vcc
; VI-NEXT:    v_bfe_u32 v46, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v46, vcc, v46, v5
; VI-NEXT:    v_add_u32_e32 v46, vcc, s6, v46
; VI-NEXT:    v_or_b32_e32 v47, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v46, v47, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v46, 16, v6
; VI-NEXT:    v_add_f32_e32 v46, 0x40c00000, v46
; VI-NEXT:    v_bfe_u32 v47, v46, 16, 1
; VI-NEXT:    v_add_u32_e32 v47, vcc, v47, v46
; VI-NEXT:    v_add_u32_e32 v47, vcc, s6, v47
; VI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; VI-NEXT:    v_or_b32_e32 v56, 0x400000, v46
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v46, v46
; VI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; VI-NEXT:    v_cndmask_b32_e32 v46, v47, v56, vcc
; VI-NEXT:    v_bfe_u32 v47, v6, 16, 1
; VI-NEXT:    v_add_u32_e32 v47, vcc, v47, v6
; VI-NEXT:    v_add_u32_e32 v47, vcc, s6, v47
; VI-NEXT:    v_or_b32_e32 v56, 0x400000, v6
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; VI-NEXT:    v_cndmask_b32_e32 v6, v47, v56, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v47, 16, v7
; VI-NEXT:    v_add_f32_e32 v47, 0x40c00000, v47
; VI-NEXT:    v_bfe_u32 v56, v47, 16, 1
; VI-NEXT:    v_add_u32_e32 v56, vcc, v56, v47
; VI-NEXT:    v_add_u32_e32 v56, vcc, s6, v56
; VI-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; VI-NEXT:    v_or_b32_e32 v57, 0x400000, v47
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v47, v47
; VI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; VI-NEXT:    v_cndmask_b32_e32 v47, v56, v57, vcc
; VI-NEXT:    v_bfe_u32 v56, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v56, vcc, v56, v7
; VI-NEXT:    v_add_u32_e32 v56, vcc, s6, v56
; VI-NEXT:    v_or_b32_e32 v57, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v56, v57, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v56, 16, v8
; VI-NEXT:    v_add_f32_e32 v56, 0x40c00000, v56
; VI-NEXT:    v_bfe_u32 v57, v56, 16, 1
; VI-NEXT:    v_add_u32_e32 v57, vcc, v57, v56
; VI-NEXT:    v_add_u32_e32 v57, vcc, s6, v57
; VI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; VI-NEXT:    v_or_b32_e32 v58, 0x400000, v56
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v56, v56
; VI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; VI-NEXT:    v_cndmask_b32_e32 v56, v57, v58, vcc
; VI-NEXT:    v_bfe_u32 v57, v8, 16, 1
; VI-NEXT:    v_add_u32_e32 v57, vcc, v57, v8
; VI-NEXT:    v_add_u32_e32 v57, vcc, s6, v57
; VI-NEXT:    v_or_b32_e32 v58, 0x400000, v8
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; VI-NEXT:    v_cndmask_b32_e32 v8, v57, v58, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v57, 16, v9
; VI-NEXT:    v_add_f32_e32 v57, 0x40c00000, v57
; VI-NEXT:    v_bfe_u32 v58, v57, 16, 1
; VI-NEXT:    v_add_u32_e32 v58, vcc, v58, v57
; VI-NEXT:    v_add_u32_e32 v58, vcc, s6, v58
; VI-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; VI-NEXT:    v_or_b32_e32 v59, 0x400000, v57
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v57, v57
; VI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; VI-NEXT:    v_cndmask_b32_e32 v57, v58, v59, vcc
; VI-NEXT:    v_bfe_u32 v58, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v58, vcc, v58, v9
; VI-NEXT:    v_add_u32_e32 v58, vcc, s6, v58
; VI-NEXT:    v_or_b32_e32 v59, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v58, v59, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v58, 16, v10
; VI-NEXT:    v_add_f32_e32 v58, 0x40c00000, v58
; VI-NEXT:    v_bfe_u32 v59, v58, 16, 1
; VI-NEXT:    v_add_u32_e32 v59, vcc, v59, v58
; VI-NEXT:    v_add_u32_e32 v59, vcc, s6, v59
; VI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; VI-NEXT:    v_or_b32_e32 v60, 0x400000, v58
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v58, v58
; VI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; VI-NEXT:    v_cndmask_b32_e32 v58, v59, v60, vcc
; VI-NEXT:    v_bfe_u32 v59, v10, 16, 1
; VI-NEXT:    v_add_u32_e32 v59, vcc, v59, v10
; VI-NEXT:    v_add_u32_e32 v59, vcc, s6, v59
; VI-NEXT:    v_or_b32_e32 v60, 0x400000, v10
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; VI-NEXT:    v_cndmask_b32_e32 v10, v59, v60, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
; VI-NEXT:    v_add_f32_e32 v59, 0x40c00000, v59
; VI-NEXT:    v_bfe_u32 v60, v59, 16, 1
; VI-NEXT:    v_add_u32_e32 v60, vcc, v60, v59
; VI-NEXT:    v_add_u32_e32 v60, vcc, s6, v60
; VI-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; VI-NEXT:    v_or_b32_e32 v61, 0x400000, v59
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v59, v59
; VI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; VI-NEXT:    v_cndmask_b32_e32 v59, v60, v61, vcc
; VI-NEXT:    v_bfe_u32 v60, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v60, vcc, v60, v11
; VI-NEXT:    v_add_u32_e32 v60, vcc, s6, v60
; VI-NEXT:    v_or_b32_e32 v61, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v60, v61, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v60, 16, v12
; VI-NEXT:    v_add_f32_e32 v60, 0x40c00000, v60
; VI-NEXT:    v_bfe_u32 v61, v60, 16, 1
; VI-NEXT:    v_add_u32_e32 v61, vcc, v61, v60
; VI-NEXT:    v_add_u32_e32 v61, vcc, s6, v61
; VI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; VI-NEXT:    v_or_b32_e32 v62, 0x400000, v60
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v60, v60
; VI-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; VI-NEXT:    v_cndmask_b32_e32 v60, v61, v62, vcc
; VI-NEXT:    v_bfe_u32 v61, v12, 16, 1
; VI-NEXT:    v_add_u32_e32 v61, vcc, v61, v12
; VI-NEXT:    v_add_u32_e32 v61, vcc, s6, v61
; VI-NEXT:    v_or_b32_e32 v62, 0x400000, v12
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; VI-NEXT:    v_cndmask_b32_e32 v12, v61, v62, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v61, 16, v13
; VI-NEXT:    v_add_f32_e32 v61, 0x40c00000, v61
; VI-NEXT:    v_bfe_u32 v62, v61, 16, 1
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v61
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v61
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v61, v61
; VI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; VI-NEXT:    v_cndmask_b32_e32 v61, v62, v63, vcc
; VI-NEXT:    v_bfe_u32 v62, v13, 16, 1
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v13
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v62, v63, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v62, 16, v14
; VI-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; VI-NEXT:    v_bfe_u32 v63, v62, 16, 1
; VI-NEXT:    v_add_u32_e32 v63, vcc, v63, v62
; VI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; VI-NEXT:    v_add_u32_e32 v63, vcc, s6, v63
; VI-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; VI-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; VI-NEXT:    v_bfe_u32 v62, v14, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v1, v63, v0, vcc
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v14
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v14
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; VI-NEXT:    v_cndmask_b32_e32 v14, v62, v63, vcc
; VI-NEXT:    v_lshlrev_b32_e32 v62, 16, v15
; VI-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; VI-NEXT:    v_bfe_u32 v63, v62, 16, 1
; VI-NEXT:    v_add_u32_e32 v63, vcc, v63, v62
; VI-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; VI-NEXT:    v_add_u32_e32 v63, vcc, s6, v63
; VI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; VI-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; VI-NEXT:    v_bfe_u32 v62, v15, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v0, v63, v0, vcc
; VI-NEXT:    v_add_u32_e32 v62, vcc, v62, v15
; VI-NEXT:    v_add_u32_e32 v62, vcc, s6, v62
; VI-NEXT:    v_or_b32_e32 v63, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v62, v63, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; VI-NEXT:    v_alignbit_b32 v15, v15, v0, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v14
; VI-NEXT:    v_alignbit_b32 v14, v0, v1, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v13
; VI-NEXT:    v_alignbit_b32 v13, v0, v61, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v12
; VI-NEXT:    v_alignbit_b32 v12, v0, v60, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v11
; VI-NEXT:    v_alignbit_b32 v11, v0, v59, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v10
; VI-NEXT:    v_alignbit_b32 v10, v0, v58, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v9
; VI-NEXT:    v_alignbit_b32 v9, v0, v57, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v8
; VI-NEXT:    v_alignbit_b32 v8, v0, v56, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v7
; VI-NEXT:    v_alignbit_b32 v7, v0, v47, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v6
; VI-NEXT:    v_alignbit_b32 v6, v0, v46, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v5
; VI-NEXT:    v_alignbit_b32 v5, v0, v45, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v4
; VI-NEXT:    v_alignbit_b32 v4, v0, v44, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v3
; VI-NEXT:    v_alignbit_b32 v3, v0, v43, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; VI-NEXT:    v_alignbit_b32 v2, v0, v42, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v31
; VI-NEXT:    v_alignbit_b32 v31, v16, v55, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v30
; VI-NEXT:    v_alignbit_b32 v30, v16, v54, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v29
; VI-NEXT:    v_alignbit_b32 v29, v16, v53, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v28
; VI-NEXT:    v_alignbit_b32 v28, v16, v52, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v27
; VI-NEXT:    v_alignbit_b32 v27, v16, v51, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v26
; VI-NEXT:    v_alignbit_b32 v26, v16, v50, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v25
; VI-NEXT:    v_alignbit_b32 v25, v16, v49, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v24
; VI-NEXT:    v_alignbit_b32 v24, v16, v48, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v23
; VI-NEXT:    v_alignbit_b32 v23, v16, v39, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v22
; VI-NEXT:    v_alignbit_b32 v22, v16, v38, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v21
; VI-NEXT:    v_alignbit_b32 v21, v16, v37, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v20
; VI-NEXT:    v_alignbit_b32 v20, v16, v36, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v19
; VI-NEXT:    v_alignbit_b32 v19, v16, v35, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v18
; VI-NEXT:    v_alignbit_b32 v18, v16, v34, 16
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v32
; VI-NEXT:    v_alignbit_b32 v1, v0, v41, 16
; VI-NEXT:    v_lshrrev_b32_e32 v0, 16, v17
; VI-NEXT:    v_alignbit_b32 v17, v16, v33, 16
; VI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; VI-NEXT:    v_alignbit_b32 v0, v0, v40, 16
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_alignbit_b32 v16, v16, v32, 16
; VI-NEXT:  .LBB104_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; VI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v64i16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_nop 0
; GFX9-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; GFX9-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB104_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX9-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX9-NEXT:    s_movk_i32 s6, 0x7fff
; GFX9-NEXT:    v_bfe_u32 v33, v32, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX9-NEXT:    v_add3_u32 v33, v33, v32, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v32
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v32, v32
; GFX9-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v33, v34, vcc
; GFX9-NEXT:    v_bfe_u32 v33, v16, 16, 1
; GFX9-NEXT:    v_add3_u32 v33, v33, v16, s6
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v33, v34, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; GFX9-NEXT:    v_add_f32_e32 v33, 0x40c00000, v33
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX9-NEXT:    v_add3_u32 v34, v34, v33, s6
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v17, 0x40c00000, v17
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v17, 16, 1
; GFX9-NEXT:    v_add3_u32 v34, v34, v17, s6
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v32, v34, v35, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v34, 16, v18
; GFX9-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX9-NEXT:    v_bfe_u32 v35, v34, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX9-NEXT:    v_add3_u32 v35, v35, v34, s6
; GFX9-NEXT:    v_or_b32_e32 v36, 0x400000, v34
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v34, v34
; GFX9-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v34, v35, v36, vcc
; GFX9-NEXT:    v_bfe_u32 v35, v18, 16, 1
; GFX9-NEXT:    v_add3_u32 v35, v35, v18, s6
; GFX9-NEXT:    v_or_b32_e32 v36, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v35, v36, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v35, 16, v19
; GFX9-NEXT:    v_add_f32_e32 v35, 0x40c00000, v35
; GFX9-NEXT:    v_bfe_u32 v36, v35, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX9-NEXT:    v_add3_u32 v36, v36, v35, s6
; GFX9-NEXT:    v_or_b32_e32 v37, 0x400000, v35
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v35, v35
; GFX9-NEXT:    v_add_f32_e32 v19, 0x40c00000, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v35, v36, v37, vcc
; GFX9-NEXT:    v_bfe_u32 v36, v19, 16, 1
; GFX9-NEXT:    v_add3_u32 v36, v36, v19, s6
; GFX9-NEXT:    v_or_b32_e32 v37, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v36, v37, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v36, 16, v20
; GFX9-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
; GFX9-NEXT:    v_bfe_u32 v37, v36, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX9-NEXT:    v_add3_u32 v37, v37, v36, s6
; GFX9-NEXT:    v_or_b32_e32 v38, 0x400000, v36
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v36, v36
; GFX9-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v36, v37, v38, vcc
; GFX9-NEXT:    v_bfe_u32 v37, v20, 16, 1
; GFX9-NEXT:    v_add3_u32 v37, v37, v20, s6
; GFX9-NEXT:    v_or_b32_e32 v38, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v37, v38, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v37, 16, v21
; GFX9-NEXT:    v_add_f32_e32 v37, 0x40c00000, v37
; GFX9-NEXT:    v_bfe_u32 v38, v37, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX9-NEXT:    v_add3_u32 v38, v38, v37, s6
; GFX9-NEXT:    v_or_b32_e32 v39, 0x400000, v37
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v37, v37
; GFX9-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v37, v38, v39, vcc
; GFX9-NEXT:    v_bfe_u32 v38, v21, 16, 1
; GFX9-NEXT:    v_add3_u32 v38, v38, v21, s6
; GFX9-NEXT:    v_or_b32_e32 v39, 0x400000, v21
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; GFX9-NEXT:    v_cndmask_b32_e32 v21, v38, v39, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v38, 16, v22
; GFX9-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; GFX9-NEXT:    v_bfe_u32 v39, v38, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX9-NEXT:    v_add3_u32 v39, v39, v38, s6
; GFX9-NEXT:    v_or_b32_e32 v48, 0x400000, v38
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v38, v38
; GFX9-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v38, v39, v48, vcc
; GFX9-NEXT:    v_bfe_u32 v39, v22, 16, 1
; GFX9-NEXT:    v_add3_u32 v39, v39, v22, s6
; GFX9-NEXT:    v_or_b32_e32 v48, 0x400000, v22
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v22, v22
; GFX9-NEXT:    v_cndmask_b32_e32 v22, v39, v48, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v39, 16, v23
; GFX9-NEXT:    v_add_f32_e32 v39, 0x40c00000, v39
; GFX9-NEXT:    v_bfe_u32 v48, v39, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX9-NEXT:    v_add3_u32 v48, v48, v39, s6
; GFX9-NEXT:    v_or_b32_e32 v49, 0x400000, v39
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v39, v39
; GFX9-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v39, v48, v49, vcc
; GFX9-NEXT:    v_bfe_u32 v48, v23, 16, 1
; GFX9-NEXT:    v_add3_u32 v48, v48, v23, s6
; GFX9-NEXT:    v_or_b32_e32 v49, 0x400000, v23
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; GFX9-NEXT:    v_cndmask_b32_e32 v23, v48, v49, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v48, 16, v24
; GFX9-NEXT:    v_add_f32_e32 v48, 0x40c00000, v48
; GFX9-NEXT:    v_bfe_u32 v49, v48, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX9-NEXT:    v_add3_u32 v49, v49, v48, s6
; GFX9-NEXT:    v_or_b32_e32 v50, 0x400000, v48
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v48, v48
; GFX9-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v48, v49, v50, vcc
; GFX9-NEXT:    v_bfe_u32 v49, v24, 16, 1
; GFX9-NEXT:    v_add3_u32 v49, v49, v24, s6
; GFX9-NEXT:    v_or_b32_e32 v50, 0x400000, v24
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v24, v24
; GFX9-NEXT:    v_cndmask_b32_e32 v24, v49, v50, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v49, 16, v25
; GFX9-NEXT:    v_add_f32_e32 v49, 0x40c00000, v49
; GFX9-NEXT:    v_bfe_u32 v50, v49, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX9-NEXT:    v_add3_u32 v50, v50, v49, s6
; GFX9-NEXT:    v_or_b32_e32 v51, 0x400000, v49
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v49, v49
; GFX9-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v49, v50, v51, vcc
; GFX9-NEXT:    v_bfe_u32 v50, v25, 16, 1
; GFX9-NEXT:    v_add3_u32 v50, v50, v25, s6
; GFX9-NEXT:    v_or_b32_e32 v51, 0x400000, v25
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; GFX9-NEXT:    v_cndmask_b32_e32 v25, v50, v51, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v50, 16, v26
; GFX9-NEXT:    v_add_f32_e32 v50, 0x40c00000, v50
; GFX9-NEXT:    v_bfe_u32 v51, v50, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX9-NEXT:    v_add3_u32 v51, v51, v50, s6
; GFX9-NEXT:    v_or_b32_e32 v52, 0x400000, v50
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v50, v50
; GFX9-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v50, v51, v52, vcc
; GFX9-NEXT:    v_bfe_u32 v51, v26, 16, 1
; GFX9-NEXT:    v_add3_u32 v51, v51, v26, s6
; GFX9-NEXT:    v_or_b32_e32 v52, 0x400000, v26
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v26, v26
; GFX9-NEXT:    v_cndmask_b32_e32 v26, v51, v52, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v51, 16, v27
; GFX9-NEXT:    v_add_f32_e32 v51, 0x40c00000, v51
; GFX9-NEXT:    v_bfe_u32 v52, v51, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX9-NEXT:    v_add3_u32 v52, v52, v51, s6
; GFX9-NEXT:    v_or_b32_e32 v53, 0x400000, v51
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v51, v51
; GFX9-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v51, v52, v53, vcc
; GFX9-NEXT:    v_bfe_u32 v52, v27, 16, 1
; GFX9-NEXT:    v_add3_u32 v52, v52, v27, s6
; GFX9-NEXT:    v_or_b32_e32 v53, 0x400000, v27
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; GFX9-NEXT:    v_cndmask_b32_e32 v27, v52, v53, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v52, 16, v28
; GFX9-NEXT:    v_add_f32_e32 v52, 0x40c00000, v52
; GFX9-NEXT:    v_bfe_u32 v53, v52, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX9-NEXT:    v_add3_u32 v53, v53, v52, s6
; GFX9-NEXT:    v_or_b32_e32 v54, 0x400000, v52
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v52, v52
; GFX9-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v52, v53, v54, vcc
; GFX9-NEXT:    v_bfe_u32 v53, v28, 16, 1
; GFX9-NEXT:    v_add3_u32 v53, v53, v28, s6
; GFX9-NEXT:    v_or_b32_e32 v54, 0x400000, v28
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v28, v28
; GFX9-NEXT:    v_cndmask_b32_e32 v28, v53, v54, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v53, 16, v29
; GFX9-NEXT:    v_add_f32_e32 v53, 0x40c00000, v53
; GFX9-NEXT:    v_bfe_u32 v54, v53, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX9-NEXT:    v_add3_u32 v54, v54, v53, s6
; GFX9-NEXT:    v_or_b32_e32 v55, 0x400000, v53
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v53, v53
; GFX9-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v53, v54, v55, vcc
; GFX9-NEXT:    v_bfe_u32 v54, v29, 16, 1
; GFX9-NEXT:    v_add3_u32 v54, v54, v29, s6
; GFX9-NEXT:    v_or_b32_e32 v55, 0x400000, v29
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; GFX9-NEXT:    v_cndmask_b32_e32 v29, v54, v55, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v54, 16, v30
; GFX9-NEXT:    v_add_f32_e32 v54, 0x40c00000, v54
; GFX9-NEXT:    v_bfe_u32 v55, v54, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX9-NEXT:    v_add3_u32 v55, v55, v54, s6
; GFX9-NEXT:    v_or_b32_e32 v40, 0x400000, v54
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v54, v54
; GFX9-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v54, v55, v40, vcc
; GFX9-NEXT:    v_bfe_u32 v55, v30, 16, 1
; GFX9-NEXT:    v_add3_u32 v55, v55, v30, s6
; GFX9-NEXT:    v_or_b32_e32 v40, 0x400000, v30
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v30, v30
; GFX9-NEXT:    v_cndmask_b32_e32 v30, v55, v40, vcc
; GFX9-NEXT:    s_waitcnt vmcnt(17)
; GFX9-NEXT:    v_lshlrev_b32_e32 v55, 16, v31
; GFX9-NEXT:    v_add_f32_e32 v55, 0x40c00000, v55
; GFX9-NEXT:    v_bfe_u32 v40, v55, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX9-NEXT:    v_add3_u32 v40, v40, v55, s6
; GFX9-NEXT:    v_or_b32_e32 v41, 0x400000, v55
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v55, v55
; GFX9-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v55, v40, v41, vcc
; GFX9-NEXT:    v_bfe_u32 v40, v31, 16, 1
; GFX9-NEXT:    v_add3_u32 v40, v40, v31, s6
; GFX9-NEXT:    v_or_b32_e32 v41, 0x400000, v31
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; GFX9-NEXT:    v_cndmask_b32_e32 v31, v40, v41, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v40, 16, v0
; GFX9-NEXT:    v_add_f32_e32 v40, 0x40c00000, v40
; GFX9-NEXT:    v_bfe_u32 v41, v40, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX9-NEXT:    v_add3_u32 v41, v41, v40, s6
; GFX9-NEXT:    v_or_b32_e32 v42, 0x400000, v40
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v40, v40
; GFX9-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v40, v41, v42, vcc
; GFX9-NEXT:    v_bfe_u32 v41, v0, 16, 1
; GFX9-NEXT:    v_add3_u32 v41, v41, v0, s6
; GFX9-NEXT:    v_or_b32_e32 v42, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v41, v42, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v41, 16, v1
; GFX9-NEXT:    v_add_f32_e32 v41, 0x40c00000, v41
; GFX9-NEXT:    v_bfe_u32 v42, v41, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX9-NEXT:    v_add3_u32 v42, v42, v41, s6
; GFX9-NEXT:    v_or_b32_e32 v43, 0x400000, v41
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v41, v41
; GFX9-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v41, v42, v43, vcc
; GFX9-NEXT:    v_bfe_u32 v42, v1, 16, 1
; GFX9-NEXT:    v_add3_u32 v42, v42, v1, s6
; GFX9-NEXT:    v_or_b32_e32 v43, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v42, v43, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v42, 16, v2
; GFX9-NEXT:    v_add_f32_e32 v42, 0x40c00000, v42
; GFX9-NEXT:    v_bfe_u32 v43, v42, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX9-NEXT:    v_add3_u32 v43, v43, v42, s6
; GFX9-NEXT:    v_or_b32_e32 v44, 0x400000, v42
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v42, v42
; GFX9-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v42, v43, v44, vcc
; GFX9-NEXT:    v_bfe_u32 v43, v2, 16, 1
; GFX9-NEXT:    v_add3_u32 v43, v43, v2, s6
; GFX9-NEXT:    v_or_b32_e32 v44, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v43, v44, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v43, 16, v3
; GFX9-NEXT:    v_add_f32_e32 v43, 0x40c00000, v43
; GFX9-NEXT:    v_bfe_u32 v44, v43, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX9-NEXT:    v_add3_u32 v44, v44, v43, s6
; GFX9-NEXT:    v_or_b32_e32 v45, 0x400000, v43
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v43, v43
; GFX9-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v43, v44, v45, vcc
; GFX9-NEXT:    v_bfe_u32 v44, v3, 16, 1
; GFX9-NEXT:    v_add3_u32 v44, v44, v3, s6
; GFX9-NEXT:    v_or_b32_e32 v45, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v44, v45, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v44, 16, v4
; GFX9-NEXT:    v_add_f32_e32 v44, 0x40c00000, v44
; GFX9-NEXT:    v_bfe_u32 v45, v44, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX9-NEXT:    v_add3_u32 v45, v45, v44, s6
; GFX9-NEXT:    v_or_b32_e32 v46, 0x400000, v44
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v44, v44
; GFX9-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v44, v45, v46, vcc
; GFX9-NEXT:    v_bfe_u32 v45, v4, 16, 1
; GFX9-NEXT:    v_add3_u32 v45, v45, v4, s6
; GFX9-NEXT:    v_or_b32_e32 v46, 0x400000, v4
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v4, v4
; GFX9-NEXT:    v_cndmask_b32_e32 v4, v45, v46, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v45, 16, v5
; GFX9-NEXT:    v_add_f32_e32 v45, 0x40c00000, v45
; GFX9-NEXT:    v_bfe_u32 v46, v45, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX9-NEXT:    v_add3_u32 v46, v46, v45, s6
; GFX9-NEXT:    v_or_b32_e32 v47, 0x400000, v45
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v45, v45
; GFX9-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v45, v46, v47, vcc
; GFX9-NEXT:    v_bfe_u32 v46, v5, 16, 1
; GFX9-NEXT:    v_add3_u32 v46, v46, v5, s6
; GFX9-NEXT:    v_or_b32_e32 v47, 0x400000, v5
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; GFX9-NEXT:    v_cndmask_b32_e32 v5, v46, v47, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v46, 16, v6
; GFX9-NEXT:    v_add_f32_e32 v46, 0x40c00000, v46
; GFX9-NEXT:    v_bfe_u32 v47, v46, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v6, 0xffff0000, v6
; GFX9-NEXT:    v_add3_u32 v47, v47, v46, s6
; GFX9-NEXT:    v_or_b32_e32 v56, 0x400000, v46
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v46, v46
; GFX9-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v46, v47, v56, vcc
; GFX9-NEXT:    v_bfe_u32 v47, v6, 16, 1
; GFX9-NEXT:    v_add3_u32 v47, v47, v6, s6
; GFX9-NEXT:    v_or_b32_e32 v56, 0x400000, v6
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v6, v6
; GFX9-NEXT:    v_cndmask_b32_e32 v6, v47, v56, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v47, 16, v7
; GFX9-NEXT:    v_add_f32_e32 v47, 0x40c00000, v47
; GFX9-NEXT:    v_bfe_u32 v56, v47, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX9-NEXT:    v_add3_u32 v56, v56, v47, s6
; GFX9-NEXT:    v_or_b32_e32 v57, 0x400000, v47
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v47, v47
; GFX9-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v47, v56, v57, vcc
; GFX9-NEXT:    v_bfe_u32 v56, v7, 16, 1
; GFX9-NEXT:    v_add3_u32 v56, v56, v7, s6
; GFX9-NEXT:    v_or_b32_e32 v57, 0x400000, v7
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; GFX9-NEXT:    v_cndmask_b32_e32 v7, v56, v57, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v56, 16, v8
; GFX9-NEXT:    v_add_f32_e32 v56, 0x40c00000, v56
; GFX9-NEXT:    v_bfe_u32 v57, v56, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX9-NEXT:    v_add3_u32 v57, v57, v56, s6
; GFX9-NEXT:    v_or_b32_e32 v58, 0x400000, v56
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v56, v56
; GFX9-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v56, v57, v58, vcc
; GFX9-NEXT:    v_bfe_u32 v57, v8, 16, 1
; GFX9-NEXT:    v_add3_u32 v57, v57, v8, s6
; GFX9-NEXT:    v_or_b32_e32 v58, 0x400000, v8
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v8, v8
; GFX9-NEXT:    v_cndmask_b32_e32 v8, v57, v58, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v57, 16, v9
; GFX9-NEXT:    v_add_f32_e32 v57, 0x40c00000, v57
; GFX9-NEXT:    v_bfe_u32 v58, v57, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX9-NEXT:    v_add3_u32 v58, v58, v57, s6
; GFX9-NEXT:    v_or_b32_e32 v59, 0x400000, v57
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v57, v57
; GFX9-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v57, v58, v59, vcc
; GFX9-NEXT:    v_bfe_u32 v58, v9, 16, 1
; GFX9-NEXT:    v_add3_u32 v58, v58, v9, s6
; GFX9-NEXT:    v_or_b32_e32 v59, 0x400000, v9
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; GFX9-NEXT:    v_cndmask_b32_e32 v9, v58, v59, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v58, 16, v10
; GFX9-NEXT:    v_add_f32_e32 v58, 0x40c00000, v58
; GFX9-NEXT:    v_bfe_u32 v59, v58, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX9-NEXT:    v_add3_u32 v59, v59, v58, s6
; GFX9-NEXT:    v_or_b32_e32 v60, 0x400000, v58
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v58, v58
; GFX9-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v58, v59, v60, vcc
; GFX9-NEXT:    v_bfe_u32 v59, v10, 16, 1
; GFX9-NEXT:    v_add3_u32 v59, v59, v10, s6
; GFX9-NEXT:    v_or_b32_e32 v60, 0x400000, v10
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v10, v10
; GFX9-NEXT:    v_cndmask_b32_e32 v10, v59, v60, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v59, 16, v11
; GFX9-NEXT:    v_add_f32_e32 v59, 0x40c00000, v59
; GFX9-NEXT:    v_bfe_u32 v60, v59, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX9-NEXT:    v_add3_u32 v60, v60, v59, s6
; GFX9-NEXT:    v_or_b32_e32 v61, 0x400000, v59
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v59, v59
; GFX9-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v59, v60, v61, vcc
; GFX9-NEXT:    v_bfe_u32 v60, v11, 16, 1
; GFX9-NEXT:    v_add3_u32 v60, v60, v11, s6
; GFX9-NEXT:    v_or_b32_e32 v61, 0x400000, v11
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; GFX9-NEXT:    v_cndmask_b32_e32 v11, v60, v61, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v60, 16, v12
; GFX9-NEXT:    v_add_f32_e32 v60, 0x40c00000, v60
; GFX9-NEXT:    v_bfe_u32 v61, v60, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX9-NEXT:    v_add3_u32 v61, v61, v60, s6
; GFX9-NEXT:    v_or_b32_e32 v62, 0x400000, v60
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v60, v60
; GFX9-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v60, v61, v62, vcc
; GFX9-NEXT:    v_bfe_u32 v61, v12, 16, 1
; GFX9-NEXT:    v_add3_u32 v61, v61, v12, s6
; GFX9-NEXT:    v_or_b32_e32 v62, 0x400000, v12
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v12, v12
; GFX9-NEXT:    v_cndmask_b32_e32 v12, v61, v62, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v61, 16, v13
; GFX9-NEXT:    v_add_f32_e32 v61, 0x40c00000, v61
; GFX9-NEXT:    v_bfe_u32 v62, v61, 16, 1
; GFX9-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX9-NEXT:    v_add3_u32 v62, v62, v61, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v61
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v61, v61
; GFX9-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v61, v62, v63, vcc
; GFX9-NEXT:    v_bfe_u32 v62, v13, 16, 1
; GFX9-NEXT:    v_add3_u32 v62, v62, v13, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v13
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; GFX9-NEXT:    v_cndmask_b32_e32 v13, v62, v63, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v62, 16, v14
; GFX9-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; GFX9-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX9-NEXT:    v_bfe_u32 v63, v62, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX9-NEXT:    v_add3_u32 v63, v63, v62, s6
; GFX9-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; GFX9-NEXT:    v_bfe_u32 v62, v14, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v63, v0, vcc
; GFX9-NEXT:    v_add3_u32 v62, v62, v14, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v14
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v14, v14
; GFX9-NEXT:    v_cndmask_b32_e32 v14, v62, v63, vcc
; GFX9-NEXT:    v_lshlrev_b32_e32 v62, 16, v15
; GFX9-NEXT:    v_add_f32_e32 v62, 0x40c00000, v62
; GFX9-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX9-NEXT:    v_bfe_u32 v63, v62, 16, 1
; GFX9-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX9-NEXT:    v_add3_u32 v63, v63, v62, s6
; GFX9-NEXT:    v_or_b32_e32 v0, 0x400000, v62
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v62, v62
; GFX9-NEXT:    v_bfe_u32 v62, v15, 16, 1
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v63, v0, vcc
; GFX9-NEXT:    v_add3_u32 v62, v62, v15, s6
; GFX9-NEXT:    v_or_b32_e32 v63, 0x400000, v15
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; GFX9-NEXT:    v_cndmask_b32_e32 v15, v62, v63, vcc
; GFX9-NEXT:    s_mov_b32 s6, 0x7060302
; GFX9-NEXT:    v_perm_b32 v15, v15, v0, s6
; GFX9-NEXT:    v_perm_b32 v14, v14, v1, s6
; GFX9-NEXT:    v_perm_b32 v1, v16, v41, s6
; GFX9-NEXT:    v_perm_b32 v0, v17, v40, s6
; GFX9-NEXT:    v_perm_b32 v17, v32, v33, s6
; GFX9-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; GFX9-NEXT:    v_perm_b32 v13, v13, v61, s6
; GFX9-NEXT:    v_perm_b32 v12, v12, v60, s6
; GFX9-NEXT:    v_perm_b32 v11, v11, v59, s6
; GFX9-NEXT:    v_perm_b32 v10, v10, v58, s6
; GFX9-NEXT:    v_perm_b32 v9, v9, v57, s6
; GFX9-NEXT:    v_perm_b32 v8, v8, v56, s6
; GFX9-NEXT:    v_perm_b32 v7, v7, v47, s6
; GFX9-NEXT:    v_perm_b32 v6, v6, v46, s6
; GFX9-NEXT:    v_perm_b32 v5, v5, v45, s6
; GFX9-NEXT:    v_perm_b32 v4, v4, v44, s6
; GFX9-NEXT:    v_perm_b32 v3, v3, v43, s6
; GFX9-NEXT:    v_perm_b32 v2, v2, v42, s6
; GFX9-NEXT:    v_perm_b32 v31, v31, v55, s6
; GFX9-NEXT:    v_perm_b32 v30, v30, v54, s6
; GFX9-NEXT:    v_perm_b32 v29, v29, v53, s6
; GFX9-NEXT:    v_perm_b32 v28, v28, v52, s6
; GFX9-NEXT:    v_perm_b32 v27, v27, v51, s6
; GFX9-NEXT:    v_perm_b32 v26, v26, v50, s6
; GFX9-NEXT:    v_perm_b32 v25, v25, v49, s6
; GFX9-NEXT:    v_perm_b32 v24, v24, v48, s6
; GFX9-NEXT:    v_perm_b32 v23, v23, v39, s6
; GFX9-NEXT:    v_perm_b32 v22, v22, v38, s6
; GFX9-NEXT:    v_perm_b32 v21, v21, v37, s6
; GFX9-NEXT:    v_perm_b32 v20, v20, v36, s6
; GFX9-NEXT:    v_perm_b32 v19, v19, v35, s6
; GFX9-NEXT:    v_perm_b32 v18, v18, v34, s6
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_perm_b32 v16, v32, v16, s6
; GFX9-NEXT:  .LBB104_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; GFX9-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v64i16:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_clause 0x1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-TRUE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-TRUE16-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-TRUE16-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-TRUE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-TRUE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT:    s_cbranch_execz .LBB104_2
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v50, 16, v22
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_lshlrev_b32 v52, 16, v24
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v32, 0x40c00000, v32
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v22, 0x40c00000, v22
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v33, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v32, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v36, 0x400000, v32
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v48, 0x400000, v33
; GFX11-TRUE16-NEXT:    v_add3_u32 v38, v38, v33, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v32, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v24, 0x40c00000, v24
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v54, 16, v26
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v35, v36, vcc_lo
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_lshlrev_b32 v34, 16, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v36, 16, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v35, v17, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v36
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v26, 0x40c00000, v26
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v18, 0x40c00000, v18
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 16, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v35, v17, 0x7fff
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_lshlrev_b32 v82, 16, v6
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v64, 16, v28
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v16, 0x40c00000, v16
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v8
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v66, 16, v30
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v28, 0x40c00000, v28
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v16, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v16
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v30, 0x40c00000, v30
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v37, v16, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v0
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v96, 16, v12
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v37, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v17
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_cndmask_b32 v33, v38, v48
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v34, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v17, v17
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v48, 16, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v4, 0xffff0000, v4
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v17, v35, v37 :: v_dual_and_b32 v6, 0xffff0000, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v38, v34, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v34
; GFX11-TRUE16-NEXT:    v_bfe_u32 v38, v18, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v20, 0x40c00000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v4, 0x40c00000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v6, 0x40c00000, v6
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v8, 0xffff0000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v35, v37, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v35, v38, v18, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v37, 0x400000, v18
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v18, v18
; GFX11-TRUE16-NEXT:    v_add3_u32 v38, v39, v36, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v8, 0x40c00000, v8
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v10, 0xffff0000, v10
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v35, v37, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v37, v19, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v48
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v10, 0x40c00000, v10
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v37, v19, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v35, v38, v39 :: v_dual_lshlrev_b32 v38, 16, v21
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v19
; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v36, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v19, v19
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v38
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v34.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v33.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v37, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v48, v36, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v36
; GFX11-TRUE16-NEXT:    v_bfe_u32 v48, v20, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v38, 16, 1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v35.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v32.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v37, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v37, v48, v20, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v39, 0x400000, v20
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v21, 0x40c00000, v21
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v20, v20
; GFX11-TRUE16-NEXT:    v_add3_u32 v48, v49, v38, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v37, v39, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v39, v21, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v50
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v36.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v39, v39, v21, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v37, v48, v49 :: v_dual_lshlrev_b32 v48, 16, v23
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v21
; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v38, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v21, v21
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v48, 0x40c00000, v48 :: v_dual_cndmask_b32 v21, v39, v49
; GFX11-TRUE16-NEXT:    v_add3_u32 v39, v50, v38, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v38
; GFX11-TRUE16-NEXT:    v_bfe_u32 v50, v22, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-TRUE16-NEXT:    v_bfe_u32 v51, v48, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v23, 0x40c00000, v23
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v37.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v39, v49, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v39, v50, v22, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v49, 0x400000, v22
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v22, v22
; GFX11-TRUE16-NEXT:    v_add3_u32 v50, v51, v48, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, 0x400000, v48
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v39, v49, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v49, v23, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v52
; GFX11-TRUE16-NEXT:    v_add3_u32 v49, v49, v23, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v39, v50, v51 :: v_dual_lshlrev_b32 v50, 16, v25
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, 0x400000, v23
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v52, v48, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v23, v23
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v50, 0x40c00000, v50 :: v_dual_cndmask_b32 v23, v49, v51
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v49, v52, v48, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, 0x400000, v48
; GFX11-TRUE16-NEXT:    v_bfe_u32 v52, v24, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-TRUE16-NEXT:    v_bfe_u32 v53, v50, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v25, 0x40c00000, v25
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v39.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v49, v51, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v49, v52, v24, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v51, 0x400000, v24
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v24, v24
; GFX11-TRUE16-NEXT:    v_add3_u32 v52, v53, v50, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v50
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v49, v51, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v51, v25, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v50, 0x40c00000, v54
; GFX11-TRUE16-NEXT:    v_add3_u32 v51, v51, v25, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v49, v52, v53 :: v_dual_lshlrev_b32 v52, 16, v27
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v25
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v54, v50, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v25, v25
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v52, 0x40c00000, v52
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v38.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v25, v51, v53, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v51, v54, v50, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v50
; GFX11-TRUE16-NEXT:    v_bfe_u32 v54, v26, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-TRUE16-NEXT:    v_bfe_u32 v55, v52, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v27, 0x40c00000, v27
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v49.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v50, v51, v53, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v51, v54, v26, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v53, 0x400000, v26
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v26, v26
; GFX11-TRUE16-NEXT:    v_add3_u32 v54, v55, v52, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v52
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v51, v53, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v53, v27, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v52, 0x40c00000, v64
; GFX11-TRUE16-NEXT:    v_add3_u32 v53, v53, v27, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v51, v54, v55 :: v_dual_lshlrev_b32 v54, 16, v29
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v27
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v64, v52, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v27, v27
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v54, 0x40c00000, v54
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v48.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v27, v53, v55, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v53, v64, v52, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v52
; GFX11-TRUE16-NEXT:    v_bfe_u32 v64, v28, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-TRUE16-NEXT:    v_bfe_u32 v65, v54, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v29, 0x40c00000, v29
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v51.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v52, v53, v55, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v53, v64, v28, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v55, 0x400000, v28
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v28, v28
; GFX11-TRUE16-NEXT:    v_add3_u32 v64, v65, v54, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v54
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v28, v53, v55, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v55, v29, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v54, 0x40c00000, v66
; GFX11-TRUE16-NEXT:    v_add3_u32 v55, v55, v29, 0x7fff
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v53, v64, v65 :: v_dual_lshlrev_b32 v64, 16, v31
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v29
; GFX11-TRUE16-NEXT:    v_bfe_u32 v66, v54, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v29, v29
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v64, 0x40c00000, v64 :: v_dual_cndmask_b32 v29, v55, v65
; GFX11-TRUE16-NEXT:    v_add3_u32 v55, v66, v54, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v54
; GFX11-TRUE16-NEXT:    v_bfe_u32 v66, v30, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-TRUE16-NEXT:    v_bfe_u32 v67, v64, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v31, 0x40c00000, v31
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v53.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v55, v65, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v55, v66, v30, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v65, 0x400000, v30
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v30, v30
; GFX11-TRUE16-NEXT:    v_add3_u32 v66, v67, v64, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, 0x400000, v64
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v55, v65, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v65, v31, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v64, 0x40c00000, v68
; GFX11-TRUE16-NEXT:    v_add3_u32 v65, v65, v31, 0x7fff
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v55, v66, v67 :: v_dual_lshlrev_b32 v66, 16, v1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, 0x400000, v31
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_bfe_u32 v68, v64, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v31, v31
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v66, 0x40c00000, v66 :: v_dual_cndmask_b32 v31, v65, v67
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v65, v68, v64, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, 0x400000, v64
; GFX11-TRUE16-NEXT:    v_bfe_u32 v68, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-TRUE16-NEXT:    v_bfe_u32 v69, v66, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v55.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v64, v65, v67, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v65, v68, v0, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v67, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add3_u32 v68, v69, v66, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, 0x400000, v66
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v0, v65, v67, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v65, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v67, 0x40c00000, v70
; GFX11-TRUE16-NEXT:    v_add3_u32 v65, v65, v1, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v66, v68, v69, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v68, 16, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v70, v67, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v3, 0xffff0000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v68, 0x40c00000, v68
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v50.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v1, v65, v69, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v65, v70, v67, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, 0x400000, v67
; GFX11-TRUE16-NEXT:    v_bfe_u32 v70, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX11-TRUE16-NEXT:    v_bfe_u32 v71, v68, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_add3_u32 v67, v70, v2, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v65, v65, v69, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v69, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_add3_u32 v70, v71, v68, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, 0x400000, v68
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v2, v67, v69, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v67, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v69, 0x40c00000, v80
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v65.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v67, v67, v3, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v68, v70, v71, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v70, 16, v5
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v69, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v5, 0xffff0000, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v70, 0x40c00000, v70
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v52.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v3, v67, v71, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v67, v80, v69, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, 0x400000, v69
; GFX11-TRUE16-NEXT:    v_bfe_u32 v80, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX11-TRUE16-NEXT:    v_bfe_u32 v81, v70, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v5, 0x40c00000, v5
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v68.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v69, v80, v4, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v67, v67, v71, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v71, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add3_u32 v80, v81, v70, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v70
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v4, v69, v71, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v69, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v70, v70
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v71, 0x40c00000, v82
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v67.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v69, v69, v5, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v70, v80, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v80, 16, v7
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v82, v71, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v7, 0xffff0000, v7
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v80, 0x40c00000, v80 :: v_dual_cndmask_b32 v5, v69, v81
; GFX11-TRUE16-NEXT:    v_add3_u32 v69, v82, v71, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v71
; GFX11-TRUE16-NEXT:    v_bfe_u32 v82, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v71, v71
; GFX11-TRUE16-NEXT:    v_bfe_u32 v83, v80, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v7, 0x40c00000, v7
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v70.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v71, v82, v6, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v69, v69, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v81, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add3_u32 v82, v83, v80, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v80
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v6, v71, v81, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v71, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v81, 0x40c00000, v84
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v69.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v71, v71, v7, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v80, v82, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v82, 16, v9
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v81, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v9, 0xffff0000, v9
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_dual_add_f32 v82, 0x40c00000, v82 :: v_dual_cndmask_b32 v7, v71, v83
; GFX11-TRUE16-NEXT:    v_add3_u32 v71, v84, v81, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v84, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v85, v82, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v9, 0x40c00000, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v80.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v81, v84, v8, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v71, v71, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v83, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add3_u32 v84, v85, v82, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v82
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v8, v81, v83, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v81, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v83, 0x40c00000, v86
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v71.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v81, v81, v9, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v82, v84, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v11
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v86, v83, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v11, 0xffff0000, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v84, 0x40c00000, v84
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v54.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v9, v81, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v81, v86, v83, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v83
; GFX11-TRUE16-NEXT:    v_bfe_u32 v86, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-TRUE16-NEXT:    v_bfe_u32 v87, v84, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v11, 0x40c00000, v11
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v82.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v83, v86, v10, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v81, v81, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_add3_u32 v86, v87, v84, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v87, 0x400000, v84
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v83, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v83, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v85, 0x40c00000, v96
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v81.h
; GFX11-TRUE16-NEXT:    v_add3_u32 v83, v83, v11, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v84, v86, v87, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v86, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v87, 16, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v96, v85, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v64.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v83, v86, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v83, 0x40c00000, v87
; GFX11-TRUE16-NEXT:    v_add3_u32 v86, v96, v85, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v87, 0x400000, v85
; GFX11-TRUE16-NEXT:    v_bfe_u32 v96, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX11-TRUE16-NEXT:    v_bfe_u32 v97, v83, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v98, 0x400000, v83
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v84.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v85, v86, v87, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v86, v96, v12, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v87, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_add3_u32 v96, v97, v83, 0x7fff
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v97, 16, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v99, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v86, v87, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v86, 0x40c00000, v97
; GFX11-TRUE16-NEXT:    v_add3_u32 v87, v99, v13, 0x7fff
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v85.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v83, v96, v98, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v97, v86, 16, 1
; GFX11-TRUE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v100, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v66.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v13, v87, v96, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v87, 0x40c00000, v98
; GFX11-TRUE16-NEXT:    v_add3_u32 v96, v97, v86, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v97, 0x400000, v86
; GFX11-TRUE16-NEXT:    v_bfe_u32 v98, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX11-TRUE16-NEXT:    v_bfe_u32 v99, v87, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e32 v15, 0x40c00000, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v83.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v86, v96, v97, vcc_lo
; GFX11-TRUE16-NEXT:    v_add3_u32 v97, v98, v14, 0x7fff
; GFX11-TRUE16-NEXT:    v_add3_u32 v98, v99, v87, 0x7fff
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v99, 0x400000, v87
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-TRUE16-NEXT:    v_bfe_u32 v96, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v101, 0x400000, v15
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v87, v98, v99, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_add3_u32 v96, v96, v15, 0x7fff
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v97, v100, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v86.h
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v96, v101, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v87.h
; GFX11-TRUE16-NEXT:  .LBB104_2: ; %end
; GFX11-TRUE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v64i16:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_clause 0x1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-FAKE16-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(1)
; GFX11-FAKE16-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-FAKE16-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-FAKE16-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-FAKE16-NEXT:    s_cbranch_execz .LBB104_2
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v33, 16, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v32, 16, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v69, 16, v2
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v71, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v33, 0x40c00000, v33 :: v_dual_add_f32 v32, 0x40c00000, v32
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v69, 0x40c00000, v69 :: v_dual_lshlrev_b32 v70, 16, v3
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v81, 16, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v33, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v32
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v32, v32
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v83, 16, v8
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v85, 16, v10
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v37, v33, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v16, 0xffff0000, v16
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v87, 16, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v12, 0xffff0000, v12
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v35, 16, v18
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v18, 0xffff0000, v18
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v34, 0x40c00000, v16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v16, v32, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v87, 0x40c00000, v87
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v35, 0x40c00000, v35 :: v_dual_add_f32 v2, 0x40c00000, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v34
; GFX11-FAKE16-NEXT:    v_add3_u32 v16, v16, v32, 0x7fff
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v17, 0xffff0000, v17
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v80, 16, v5
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v82, 16, v7
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v84, 16, v9
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v16, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v36, 0x40c00000, v17
; GFX11-FAKE16-NEXT:    v_bfe_u32 v17, v34, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v34, v34
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v38, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v18, 16, v19
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v86, 16, v11
; GFX11-FAKE16-NEXT:    v_add3_u32 v17, v17, v34, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v34, v36, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v17, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v17, 0x400000, v33
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v33, v33
; GFX11-FAKE16-NEXT:    v_add3_u32 v33, v34, v36, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v34, 0x400000, v36
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v37, v17, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v35, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v36, v36
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v35
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v37, v35, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v37, v38, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v18
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v35, v35
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v19, 0xffff0000, v19
; GFX11-FAKE16-NEXT:    v_perm_b32 v17, v33, v17, 0x7060302
; GFX11-FAKE16-NEXT:    v_perm_b32 v16, v32, v16, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v18, v34, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v34, v37, v38, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v36, v39, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v37, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v35, 0x400000, v38
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v38, v38
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v19, 16, v20
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v20, 0xffff0000, v20
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v37, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v34, v35, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v36, v39, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v39
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v48, 0x40c00000, v19
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v39, 0x40c00000, v20
; GFX11-FAKE16-NEXT:    v_perm_b32 v18, v34, v18, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v35, v38, v37, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v36, 0x400000, v37
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v48, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v37, v37
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v20, 16, v21
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v48
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v21, 0xffff0000, v21
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v35, v36, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v38, v48, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v39, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v49, 0x40c00000, v20 :: v_dual_add_f32 v48, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v21, 16, v22
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v22, 0xffff0000, v22
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v20, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v36, v38, v39, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v37, 0x400000, v39
; GFX11-FAKE16-NEXT:    v_bfe_u32 v38, v49, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v39, v39
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v48, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v50, 0x40c00000, v21
; GFX11-FAKE16-NEXT:    v_perm_b32 v19, v35, v19, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v36, v36, v37, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v38, v49, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v49
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v49, 0x40c00000, v22 :: v_dual_lshlrev_b32 v22, 16, v23
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v23, 0xffff0000, v23
; GFX11-FAKE16-NEXT:    v_perm_b32 v20, v36, v20, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v21, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v37, v39, v48, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v38, 0x400000, v48
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v48, v48
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v49, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v51, 0x40c00000, v22
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v37, v38, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v39, v50, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    v_perm_b32 v21, v37, v21, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v39, v50, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v50
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v50, 0x40c00000, v23
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v38, v48, v49, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v39, 0x400000, v49
; GFX11-FAKE16-NEXT:    v_bfe_u32 v48, v51, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v49, v49
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v23, 16, v24
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v50, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v24, 0xffff0000, v24
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v38, v39, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v39, v48, v51, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v51
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v52, 0x40c00000, v23 :: v_dual_add_f32 v51, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v24, 16, v25
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v25, 0xffff0000, v25
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v39, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v39, v49, v50, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v48, 0x400000, v50
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v50, v50
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v51, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v53, 0x40c00000, v24
; GFX11-FAKE16-NEXT:    v_perm_b32 v22, v38, v22, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v39, v48, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v49, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v23, v39, v23, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v49, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v52, 0x40c00000, v25
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v24, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v48, v50, v51, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v49, 0x400000, v51
; GFX11-FAKE16-NEXT:    v_bfe_u32 v50, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v51, v51
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v25, 16, v26
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v52, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v26, 0xffff0000, v26
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v48, v49, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v50, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v54, 0x40c00000, v25 :: v_dual_add_f32 v53, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v26, 16, v27
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v27, 0xffff0000, v27
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v49, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v49, v51, v52, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v50, 0x400000, v52
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v52, v52
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v53, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v55, 0x40c00000, v26
; GFX11-FAKE16-NEXT:    v_perm_b32 v24, v48, v24, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v49, v50, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v51, v54, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v25, v49, v25, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v51, v54, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v54
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v54, 0x40c00000, v27
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v50, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v50, v52, v53, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v51, 0x400000, v53
; GFX11-FAKE16-NEXT:    v_bfe_u32 v52, v55, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v53, v53
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v27, 16, v28
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v54, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v28, 0xffff0000, v28
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v50, v51, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v51, v52, v55, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v55
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v64, 0x40c00000, v27 :: v_dual_add_f32 v55, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v29, 0xffff0000, v29
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v51, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v51, v53, v54, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v52, 0x400000, v54
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v54, v54
; GFX11-FAKE16-NEXT:    v_bfe_u32 v54, v55, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v65, 0x40c00000, v28
; GFX11-FAKE16-NEXT:    v_perm_b32 v26, v50, v26, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v51, v52, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v53, v64, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v27, v51, v27, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v53, v64, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v64
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v64, 0x40c00000, v29
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v28, v52, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v52, v54, v55, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v53, 0x400000, v55
; GFX11-FAKE16-NEXT:    v_bfe_u32 v54, v65, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v55, v55
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v29, 16, v30
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v64, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v30, 0xffff0000, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v52, v53, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v54, v65, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v65
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v66, 0x40c00000, v29 :: v_dual_add_f32 v65, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v30, 16, v31
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v31, 0xffff0000, v31
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v53, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v53, v55, v64, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v54, 0x400000, v64
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v64, v64
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v65, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v67, 0x40c00000, v30
; GFX11-FAKE16-NEXT:    v_perm_b32 v28, v52, v28, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v53, v54, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v55, v66, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_perm_b32 v29, v53, v29, 0x7060302
; GFX11-FAKE16-NEXT:    v_add3_u32 v54, v55, v66, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v66
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v66, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v30, v54, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v54, v64, v65, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v55, 0x400000, v65
; GFX11-FAKE16-NEXT:    v_bfe_u32 v64, v67, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v65, v65
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v31, 16, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v65, v66, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v54, v54, v55, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v55, v64, v67, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v67
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v68, 0x40c00000, v31
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v67, 16, v1
; GFX11-FAKE16-NEXT:    v_perm_b32 v30, v54, v30, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v31, v55, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v55, v65, v66, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v64, 0x400000, v66
; GFX11-FAKE16-NEXT:    v_bfe_u32 v65, v68, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v66, v66
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v67, 0x40c00000, v67
; GFX11-FAKE16-NEXT:    v_bfe_u32 v66, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v55, v64, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v64, v65, v68, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v65, 0x400000, v68
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v68, v68
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v67, 16, 1
; GFX11-FAKE16-NEXT:    v_perm_b32 v31, v55, v31, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v64, v65, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v65, v66, v0, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v0, v65, v66 :: v_dual_and_b32 v1, 0xffff0000, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; GFX11-FAKE16-NEXT:    v_add3_u32 v65, v68, v67, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v66, 0x400000, v67
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v67, v67
; GFX11-FAKE16-NEXT:    v_perm_b32 v0, v0, v64, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v65, v66, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v66, v68, v1, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v69, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v66, v67, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v66, v68, v69, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v67, 0x400000, v69
; GFX11-FAKE16-NEXT:    v_bfe_u32 v68, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v69, v69
; GFX11-FAKE16-NEXT:    v_perm_b32 v1, v1, v65, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v66, v67, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v67, v68, v2, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v68, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v67, v68 :: v_dual_and_b32 v3, 0xffff0000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v70, 0x40c00000, v70 :: v_dual_add_f32 v3, 0x40c00000, v3
; GFX11-FAKE16-NEXT:    v_perm_b32 v2, v2, v66, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v69, v70, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v68, 0x400000, v70
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v70, v70
; GFX11-FAKE16-NEXT:    v_add3_u32 v67, v69, v70, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v69, v3, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v67, v67, v68, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v68, v69, v3, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v3, v68, v69 :: v_dual_and_b32 v4, 0xffff0000, v4
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v71, 0x40c00000, v71 :: v_dual_add_f32 v4, 0x40c00000, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v3, v3, v67, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v70, v71, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v69, 0x400000, v71
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v71, v71
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v68, v70, v71, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v70, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v68, v68, v69, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v69, v70, v4, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v70, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v4, v69, v70 :: v_dual_and_b32 v5, 0xffff0000, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v80, 0x40c00000, v80 :: v_dual_add_f32 v5, 0x40c00000, v5
; GFX11-FAKE16-NEXT:    v_perm_b32 v4, v4, v68, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v80, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v70, 0x400000, v80
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-FAKE16-NEXT:    v_add3_u32 v69, v71, v80, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v5, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v69, v70, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v70, v71, v5, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v5, v70, v71 :: v_dual_and_b32 v6, 0xffff0000, v6
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v81, 0x40c00000, v81 :: v_dual_add_f32 v6, 0x40c00000, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v5, v5, v69, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v80, v81, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v81
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v70, v80, v81, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v80, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v70, v70, v71, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v71, v80, v6, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v80, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v6, v71, v80 :: v_dual_and_b32 v7, 0xffff0000, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v82, 0x40c00000, v82 :: v_dual_add_f32 v7, 0x40c00000, v7
; GFX11-FAKE16-NEXT:    v_perm_b32 v6, v6, v70, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v81, v82, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v80, 0x400000, v82
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-FAKE16-NEXT:    v_add3_u32 v71, v81, v82, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v81, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v71, v71, v80, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v80, v81, v7, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v81, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v7, v80, v81 :: v_dual_and_b32 v8, 0xffff0000, v8
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v83, 0x40c00000, v83 :: v_dual_add_f32 v8, 0x40c00000, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v7, v7, v71, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v83, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v81, 0x400000, v83
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v80, v82, v83, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v82, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v80, v80, v81, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v81, v82, v8, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v81, v82 :: v_dual_and_b32 v9, 0xffff0000, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v84, 0x40c00000, v84 :: v_dual_add_f32 v9, 0x40c00000, v9
; GFX11-FAKE16-NEXT:    v_perm_b32 v8, v8, v80, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v84, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, 0x400000, v84
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-FAKE16-NEXT:    v_add3_u32 v81, v83, v84, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v9, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v81, v81, v82, vcc_lo
; GFX11-FAKE16-NEXT:    v_add3_u32 v82, v83, v9, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, 0x400000, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v9, v82, v83 :: v_dual_and_b32 v10, 0xffff0000, v10
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v85, 0x40c00000, v85 :: v_dual_add_f32 v10, 0x40c00000, v10
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v9, v9, v81, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v84, v85, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, 0x400000, v85
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v82, v84, v85, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v84, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v82, v82, v83, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add3_u32 v83, v84, v10, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v10, v83, v84 :: v_dual_and_b32 v11, 0xffff0000, v11
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v86, 0x40c00000, v86 :: v_dual_add_f32 v11, 0x40c00000, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v10, v10, v82, 0x7060302
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v85, v86, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, 0x400000, v86
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v86, v86
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v96, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_add3_u32 v83, v85, v86, 0x7fff
; GFX11-FAKE16-NEXT:    v_bfe_u32 v85, v11, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v83, v83, v84 :: v_dual_lshlrev_b32 v84, 16, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v86, v87, 16, 1
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v13, 0xffff0000, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v12, 0x40c00000, v12
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v97, 0x400000, v87
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v84, 0x40c00000, v84
; GFX11-FAKE16-NEXT:    v_add3_u32 v86, v86, v87, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v13, 0x40c00000, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v98, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v84, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v85, v85, v11, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v86, v86, v97, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v97, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_add3_u32 v87, v98, v12, 0x7fff
; GFX11-FAKE16-NEXT:    v_add3_u32 v98, v99, v84, 0x7fff
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v99, 16, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v14, 0xffff0000, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v100, 0x400000, v84
; GFX11-FAKE16-NEXT:    v_bfe_u32 v101, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v87, v97 :: v_dual_add_f32 v87, 0x40c00000, v99
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-FAKE16-NEXT:    v_add_f32_e32 v14, 0x40c00000, v14
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add3_u32 v97, v101, v13, 0x7fff
; GFX11-FAKE16-NEXT:    v_perm_b32 v12, v12, v86, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v87, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v84, v98, v100, vcc_lo
; GFX11-FAKE16-NEXT:    v_lshlrev_b32_e32 v98, 16, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v101, 0x400000, v87
; GFX11-FAKE16-NEXT:    v_bfe_u32 v102, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_add3_u32 v99, v99, v87, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v100, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_add_f32 v98, 0x40c00000, v98 :: v_dual_cndmask_b32 v87, v99, v101
; GFX11-FAKE16-NEXT:    v_add3_u32 v101, v102, v14, 0x7fff
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v102, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_and_b32_e32 v15, 0xffff0000, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v103, v98, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v112, 0x400000, v98
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v14, v101, v102 :: v_dual_add_f32 v15, 0x40c00000, v15
; GFX11-FAKE16-NEXT:    v_add3_u32 v103, v103, v98, 0x7fff
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_perm_b32 v14, v14, v87, 0x7060302
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v113, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v98, v103, v112, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_add3_u32 v99, v99, v15, 0x7fff
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v15, v99, v113, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_perm_b32 v15, v15, v98, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v97, v100, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_perm_b32 v13, v13, v84, 0x7060302
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v85, v96, vcc_lo
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_1)
; GFX11-FAKE16-NEXT:    v_perm_b32 v11, v11, v83, 0x7060302
; GFX11-FAKE16-NEXT:  .LBB104_2: ; %end
; GFX11-FAKE16-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define inreg <64 x i16> @bitcast_v64bf16_to_v64i16_scalar(<64 x bfloat> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64bf16_to_v64i16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:508 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    s_and_b32 s14, s24, 0xffff0000
; SI-NEXT:    s_lshl_b32 s15, s24, 16
; SI-NEXT:    s_and_b32 s24, s23, 0xffff0000
; SI-NEXT:    s_lshl_b32 s23, s23, 16
; SI-NEXT:    s_and_b32 s30, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v1
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    s_lshl_b32 s13, s25, 16
; SI-NEXT:    v_readfirstlane_b32 s42, v17
; SI-NEXT:    v_readfirstlane_b32 s44, v16
; SI-NEXT:    v_readfirstlane_b32 s46, v15
; SI-NEXT:    v_readfirstlane_b32 s56, v14
; SI-NEXT:    v_readfirstlane_b32 s58, v13
; SI-NEXT:    v_readfirstlane_b32 s60, v12
; SI-NEXT:    v_readfirstlane_b32 s62, v11
; SI-NEXT:    v_readfirstlane_b32 s72, v10
; SI-NEXT:    v_readfirstlane_b32 s74, v9
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    v_readfirstlane_b32 s78, v7
; SI-NEXT:    v_readfirstlane_b32 s88, v6
; SI-NEXT:    v_readfirstlane_b32 s90, v5
; SI-NEXT:    v_readfirstlane_b32 s92, v4
; SI-NEXT:    v_readfirstlane_b32 s94, v3
; SI-NEXT:    s_and_b32 s34, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s23
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    s_and_b32 s4, s29, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s29, 16
; SI-NEXT:    s_and_b32 s6, s28, 0xffff0000
; SI-NEXT:    s_lshl_b32 s7, s28, 16
; SI-NEXT:    s_and_b32 s8, s27, 0xffff0000
; SI-NEXT:    s_lshl_b32 s9, s27, 16
; SI-NEXT:    s_and_b32 s10, s26, 0xffff0000
; SI-NEXT:    s_lshl_b32 s11, s26, 16
; SI-NEXT:    s_and_b32 s12, s25, 0xffff0000
; SI-NEXT:    s_and_b32 s25, s22, 0xffff0000
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_and_b32 s26, s21, 0xffff0000
; SI-NEXT:    s_lshl_b32 s21, s21, 16
; SI-NEXT:    s_and_b32 s27, s20, 0xffff0000
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_and_b32 s28, s19, 0xffff0000
; SI-NEXT:    s_lshl_b32 s19, s19, 16
; SI-NEXT:    s_and_b32 s29, s18, 0xffff0000
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_and_b32 s40, s17, 0xffff0000
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_and_b32 s41, s16, 0xffff0000
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    s_and_b32 s43, s42, 0xffff0000
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_and_b32 s45, s44, 0xffff0000
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_and_b32 s57, s56, 0xffff0000
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    s_and_b32 s59, s58, 0xffff0000
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_and_b32 s61, s60, 0xffff0000
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_and_b32 s63, s62, 0xffff0000
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_and_b32 s73, s72, 0xffff0000
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    s_and_b32 s75, s74, 0xffff0000
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_and_b32 s77, s76, 0xffff0000
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_and_b32 s79, s78, 0xffff0000
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_and_b32 s89, s88, 0xffff0000
; SI-NEXT:    s_lshl_b32 s88, s88, 16
; SI-NEXT:    s_and_b32 s91, s90, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s90, 16
; SI-NEXT:    s_and_b32 s93, s92, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s92, 16
; SI-NEXT:    s_and_b32 s95, s94, 0xffff0000
; SI-NEXT:    s_lshl_b32 s94, s94, 16
; SI-NEXT:    s_and_b32 s36, vcc_lo, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, vcc_lo, 16
; SI-NEXT:    v_readfirstlane_b32 vcc_lo, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:340 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s13
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cmp_lg_u32 vcc_lo, 0
; SI-NEXT:    v_mul_f32_e64 v42, 1.0, s16
; SI-NEXT:    v_mul_f32_e64 v46, 1.0, s41
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_mul_f32_e64 v56, 1.0, s17
; SI-NEXT:    v_mul_f32_e64 v47, 1.0, s40
; SI-NEXT:    v_mul_f32_e64 v38, 1.0, s18
; SI-NEXT:    v_mul_f32_e64 v45, 1.0, s29
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_mul_f32_e64 v58, 1.0, s19
; SI-NEXT:    v_mul_f32_e64 v57, 1.0, s28
; SI-NEXT:    v_mul_f32_e64 v43, 1.0, s20
; SI-NEXT:    v_mul_f32_e64 v44, 1.0, s27
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mul_f32_e64 v61, 1.0, s21
; SI-NEXT:    v_mul_f32_e64 v59, 1.0, s26
; SI-NEXT:    v_mul_f32_e64 v55, 1.0, s22
; SI-NEXT:    v_mul_f32_e64 v40, 1.0, s25
; SI-NEXT:    v_mul_f32_e64 v32, 1.0, s24
; SI-NEXT:    v_mul_f32_e64 v53, 1.0, s15
; SI-NEXT:    v_mul_f32_e64 v54, 1.0, s14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:344 ; 4-byte Folded Spill
; SI-NEXT:    v_mul_f32_e64 v34, 1.0, s12
; SI-NEXT:    v_mul_f32_e64 v48, 1.0, s11
; SI-NEXT:    v_mul_f32_e64 v49, 1.0, s10
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s9
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s8
; SI-NEXT:    v_mul_f32_e64 v36, 1.0, s7
; SI-NEXT:    v_mul_f32_e64 v39, 1.0, s6
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s5
; SI-NEXT:    v_mul_f32_e64 v37, 1.0, s4
; SI-NEXT:    v_mul_f32_e64 v33, 1.0, s37
; SI-NEXT:    v_mul_f32_e64 v35, 1.0, s36
; SI-NEXT:    v_mul_f32_e64 v51, 1.0, s35
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s34
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s31
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s30
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s94
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s95
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s92
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s93
; SI-NEXT:    v_mul_f32_e64 v41, 1.0, s90
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s91
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s88
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s89
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s78
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s79
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s76
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s77
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s74
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s75
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s72
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s73
; SI-NEXT:    v_mul_f32_e64 v50, 1.0, s62
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s63
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s60
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s61
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s58
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s59
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s56
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s57
; SI-NEXT:    v_mul_f32_e64 v52, 1.0, s46
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s47
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s44
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s45
; SI-NEXT:    v_mul_f32_e64 v62, 1.0, s42
; SI-NEXT:    v_mul_f32_e64 v60, 1.0, s43
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:348 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:352 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:356 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:360 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:364 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:368 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:372 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:376 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:380 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:384 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v21, off, s[0:3], s32 offset:388 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:392 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:396 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v27, off, s[0:3], s32 offset:400 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:404 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:408 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:412 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 offset:416 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v36, off, s[0:3], s32 offset:420 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:424 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v48, off, s[0:3], s32 offset:428 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v49, off, s[0:3], s32 offset:432 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:436 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v54, off, s[0:3], s32 offset:440 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:444 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:448 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:452 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:456 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:460 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:464 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:468 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:472 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:476 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:480 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:484 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:488 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:492 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:496 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:500 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v37, off, s[0:3], s32 offset:504 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:328 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:332 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:336 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB105_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v42
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v46
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v47
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v28
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v45
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v31
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v38
; SI-NEXT:    buffer_store_dword v38, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v39, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v57
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v34
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    v_mov_b32_e32 v31, v38
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v61
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v52
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v43
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v22
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v55
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v56
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v14
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v44
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v59
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v40
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v32
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v48
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v36
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v53
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v18
; SI-NEXT:    v_mov_b32_e32 v48, v59
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v50
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v62
; SI-NEXT:    v_mov_b32_e32 v43, v34
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v54
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v15
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v37
; SI-NEXT:    v_mov_b32_e32 v37, v57
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v49
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v58
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v30
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v25
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v39
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v10
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v35
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v23
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v6, v25
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v7
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v51
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v4
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v41
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v60
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v33
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v0
; SI-NEXT:    v_mov_b32_e32 v34, v2
; SI-NEXT:    s_branch .LBB105_3
; SI-NEXT:  .LBB105_2:
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:340 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v43, v34
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:344 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    v_mov_b32_e32 v6, v25
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr54
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr40
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:  .LBB105_3: ; %Flow
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:328 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:332 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:336 ; 4-byte Folded Reload
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_cbranch_vccnz .LBB105_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v62
; SI-NEXT:    v_add_f32_e32 v35, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v52
; SI-NEXT:    v_add_f32_e32 v45, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v14
; SI-NEXT:    v_add_f32_e32 v27, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v50
; SI-NEXT:    v_add_f32_e32 v58, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v16
; SI-NEXT:    v_add_f32_e32 v31, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v10
; SI-NEXT:    v_add_f32_e32 v52, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v41
; SI-NEXT:    v_add_f32_e32 v54, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v4
; SI-NEXT:    v_add_f32_e32 v49, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v51
; SI-NEXT:    v_add_f32_e32 v40, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v23
; SI-NEXT:    v_add_f32_e32 v23, 0x40c00000, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v30
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v43
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:500 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v60
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v20
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v26
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v6
; SI-NEXT:    v_add_f32_e32 v47, 0x40c00000, v0
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v34
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_add_f32_e32 v60, 0x40c00000, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v33
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v3
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v18
; SI-NEXT:    v_add_f32_e32 v18, 0x40c00000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:496 ; 4-byte Folded Reload
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v12
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v8
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff0000, v29
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v3
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:504 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:492 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v42, 0x40c00000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:488 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:484 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v29, 0x40c00000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:480 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:476 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v20, 0x40c00000, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:468 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v1, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:472 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v25, 0x40c00000, v1
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v25
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:460 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:464 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v22, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v22
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:452 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:456 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v33, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v33
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:444 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:448 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v17, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v17
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:436 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:440 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v15, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v15
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:428 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:432 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v13, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v13
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:420 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:424 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v11, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v11
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:412 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:416 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v9, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v9
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:404 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:408 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v7, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v7
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:392 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:396 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v5, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v5
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:384 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:388 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v3, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v3
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:376 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:380 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v2
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v2
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v2
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:368 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:372 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v4, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v4
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:400 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:364 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v4
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v6, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v6
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:356 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:360 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v26, 0xffff0000, v6
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v8, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v8
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:348 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:352 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v8
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
; SI-NEXT:    v_add_f32_e32 v10, 0x40c00000, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x40c00000, v0
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v10
; SI-NEXT:    v_lshr_b64 v[38:39], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v1, v36
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v46
; SI-NEXT:    v_lshr_b64 v[35:36], v[35:36], 16
; SI-NEXT:    v_mov_b32_e32 v36, v60
; SI-NEXT:    v_lshr_b64 v[45:46], v[45:46], 16
; SI-NEXT:    v_mov_b32_e32 v37, v61
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v10
; SI-NEXT:    v_lshr_b64 v[61:62], v[42:43], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v28
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v1, v59
; SI-NEXT:    v_lshr_b64 v[58:59], v[58:59], 16
; SI-NEXT:    v_mov_b32_e32 v60, v48
; SI-NEXT:    v_mov_b32_e32 v59, v47
; SI-NEXT:    v_mov_b32_e32 v47, v41
; SI-NEXT:    v_mov_b32_e32 v46, v40
; SI-NEXT:    v_mov_b32_e32 v40, v32
; SI-NEXT:    v_mov_b32_e32 v39, v31
; SI-NEXT:    v_mov_b32_e32 v42, v21
; SI-NEXT:    v_mov_b32_e32 v41, v20
; SI-NEXT:    v_mov_b32_e32 v31, v30
; SI-NEXT:    v_mov_b32_e32 v30, v29
; SI-NEXT:    v_mov_b32_e32 v48, v19
; SI-NEXT:    v_and_b32_e32 v20, 0xffff0000, v22
; SI-NEXT:    v_lshr_b64 v[21:22], v[30:31], 16
; SI-NEXT:    v_lshr_b64 v[28:29], v[57:58], 16
; SI-NEXT:    v_mov_b32_e32 v56, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff0000, v2
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[1:2], v[39:40], 16
; SI-NEXT:    v_and_b32_e32 v2, 0xffff0000, v3
; SI-NEXT:    v_lshr_b64 v[3:4], v[52:53], 16
; SI-NEXT:    v_and_b32_e32 v4, 0xffff0000, v5
; SI-NEXT:    v_lshr_b64 v[5:6], v[54:55], 16
; SI-NEXT:    v_and_b32_e32 v6, 0xffff0000, v7
; SI-NEXT:    v_lshr_b64 v[7:8], v[49:50], 16
; SI-NEXT:    v_and_b32_e32 v8, 0xffff0000, v9
; SI-NEXT:    v_lshr_b64 v[9:10], v[46:47], 16
; SI-NEXT:    v_and_b32_e32 v10, 0xffff0000, v11
; SI-NEXT:    v_lshr_b64 v[11:12], v[23:24], 16
; SI-NEXT:    v_and_b32_e32 v12, 0xffff0000, v13
; SI-NEXT:    v_mov_b32_e32 v13, v60
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v13, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[13:14], v[59:60], 16
; SI-NEXT:    v_and_b32_e32 v14, 0xffff0000, v15
; SI-NEXT:    v_lshr_b64 v[15:16], v[36:37], 16
; SI-NEXT:    v_and_b32_e32 v16, 0xffff0000, v17
; SI-NEXT:    v_lshr_b64 v[17:18], v[18:19], 16
; SI-NEXT:    v_mov_b32_e32 v18, v43
; SI-NEXT:    buffer_store_dword v17, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v54, v24
; SI-NEXT:    v_and_b32_e32 v23, 0xffff0000, v25
; SI-NEXT:    v_lshr_b64 v[24:25], v[41:42], 16
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v33
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[23:24], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[20:21], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[60:61], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v46, v45
; SI-NEXT:    v_mov_b32_e32 v39, v50
; SI-NEXT:    v_lshr_b64 v[22:23], v[4:5], 16
; SI-NEXT:    v_lshr_b64 v[49:50], v[2:3], 16
; SI-NEXT:    v_lshr_b64 v[32:33], v[0:1], 16
; SI-NEXT:    v_mov_b32_e32 v59, v58
; SI-NEXT:    v_mov_b32_e32 v36, v35
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[16:17], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[14:15], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[12:13], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[10:11], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[8:9], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[6:7], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[26:27], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[44:45], 16
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[18:19], v[34:35], 16
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v19, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:  .LBB105_5: ; %end
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v1
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v3
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v5
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v61
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v7
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v9
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v49
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v42
; SI-NEXT:    v_or_b32_e32 v1, v2, v1
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v31
; SI-NEXT:    v_or_b32_e32 v3, v4, v3
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v25
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v48
; SI-NEXT:    v_or_b32_e32 v7, v8, v7
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v44
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v37
; SI-NEXT:    v_or_b32_e32 v9, v10, v9
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v25
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v11, v12, v11
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v12, v12, v14
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v54
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v14, v14, v16
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v47
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v16, v16, v18
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v39
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v18, v18, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v55
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v20, v20, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v53
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v32
; SI-NEXT:    v_or_b32_e32 v22, v22, v24
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v40
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v59
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v29
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v36
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v28
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v27
; SI-NEXT:    v_or_b32_e32 v26, v26, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v56
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v29
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v33
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v46
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v38
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:508 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64bf16_to_v64i16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v35, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v35, s30, 0
; VI-NEXT:    v_writelane_b32 v35, s31, 1
; VI-NEXT:    v_writelane_b32 v35, s36, 2
; VI-NEXT:    v_writelane_b32 v35, s37, 3
; VI-NEXT:    v_writelane_b32 v35, s38, 4
; VI-NEXT:    v_writelane_b32 v35, s39, 5
; VI-NEXT:    v_writelane_b32 v35, s48, 6
; VI-NEXT:    v_writelane_b32 v35, s49, 7
; VI-NEXT:    v_writelane_b32 v35, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v35, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB105_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB105_4
; VI-NEXT:  .LBB105_2: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x40c00000
; VI-NEXT:    s_lshl_b32 s4, s48, 16
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    s_and_b32 s4, s48, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v16, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    s_lshl_b32 s5, s28, 16
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v0
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    s_and_b32 s5, s28, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; VI-NEXT:    v_add_f32_e32 v1, s5, v32
; VI-NEXT:    v_bfe_u32 v2, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v1
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    s_lshl_b32 s5, s30, 16
; VI-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; VI-NEXT:    v_add_f32_e32 v2, s5, v32
; VI-NEXT:    v_bfe_u32 v3, v2, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v2
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; VI-NEXT:    s_and_b32 s5, s30, 0xffff0000
; VI-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; VI-NEXT:    v_add_f32_e32 v3, s5, v32
; VI-NEXT:    v_bfe_u32 v4, v3, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    v_add_u32_e32 v4, vcc, v4, v3
; VI-NEXT:    s_and_b32 s5, s26, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v4, vcc, 0x7fff, v4
; VI-NEXT:    v_lshrrev_b64 v[12:13], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s26, 16
; VI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_lshrrev_b64 v[14:15], 16, v[2:3]
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s24, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[10:11], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s24, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s22, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[8:9], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s22, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s20, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[6:7], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s20, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s18, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[4:5], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s18, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v2, vcc
; VI-NEXT:    v_bfe_u32 v2, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v2, vcc, v2, v0
; VI-NEXT:    v_add_u32_e32 v2, vcc, 0x7fff, v2
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s16, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[2:3], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v0, s5, v32
; VI-NEXT:    v_bfe_u32 v1, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v1, vcc, v1, v0
; VI-NEXT:    s_lshl_b32 s4, s16, 16
; VI-NEXT:    v_add_u32_e32 v1, vcc, 0x7fff, v1
; VI-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_add_f32_e32 v0, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v1, v1, v3, vcc
; VI-NEXT:    v_bfe_u32 v3, v0, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v0
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v0
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; VI-NEXT:    v_cndmask_b32_e32 v0, v3, v5, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; VI-NEXT:    s_and_b32 s5, s17, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[0:1], 16, v[0:1]
; VI-NEXT:    v_add_f32_e32 v1, s5, v32
; VI-NEXT:    v_bfe_u32 v3, v1, 16, 1
; VI-NEXT:    v_add_u32_e32 v3, vcc, v3, v1
; VI-NEXT:    v_add_u32_e32 v3, vcc, 0x7fff, v3
; VI-NEXT:    s_lshl_b32 s4, s17, 16
; VI-NEXT:    v_or_b32_e32 v5, 0x400000, v1
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; VI-NEXT:    v_cndmask_b32_e32 v1, v3, v5, vcc
; VI-NEXT:    v_add_f32_e32 v3, s4, v32
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_and_b32 s5, s19, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_add_f32_e32 v3, s5, v32
; VI-NEXT:    v_cndmask_b32_e32 v18, v5, v7, vcc
; VI-NEXT:    v_bfe_u32 v5, v3, 16, 1
; VI-NEXT:    v_add_u32_e32 v5, vcc, v5, v3
; VI-NEXT:    v_add_u32_e32 v5, vcc, 0x7fff, v5
; VI-NEXT:    s_lshl_b32 s4, s19, 16
; VI-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; VI-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc
; VI-NEXT:    v_add_f32_e32 v5, s4, v32
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_and_b32 s5, s21, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_add_f32_e32 v5, s5, v32
; VI-NEXT:    v_mov_b32_e32 v1, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v7, v9, vcc
; VI-NEXT:    v_bfe_u32 v7, v5, 16, 1
; VI-NEXT:    v_add_u32_e32 v7, vcc, v7, v5
; VI-NEXT:    v_add_u32_e32 v7, vcc, 0x7fff, v7
; VI-NEXT:    s_lshl_b32 s4, s21, 16
; VI-NEXT:    v_or_b32_e32 v9, 0x400000, v5
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v5, v5
; VI-NEXT:    v_cndmask_b32_e32 v5, v7, v9, vcc
; VI-NEXT:    v_add_f32_e32 v7, s4, v32
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v3
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_and_b32 s5, s23, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_add_f32_e32 v7, s5, v32
; VI-NEXT:    v_mov_b32_e32 v3, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v9, v11, vcc
; VI-NEXT:    v_bfe_u32 v9, v7, 16, 1
; VI-NEXT:    v_add_u32_e32 v9, vcc, v9, v7
; VI-NEXT:    v_add_u32_e32 v9, vcc, 0x7fff, v9
; VI-NEXT:    s_lshl_b32 s4, s23, 16
; VI-NEXT:    v_or_b32_e32 v11, 0x400000, v7
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v7, v7
; VI-NEXT:    v_cndmask_b32_e32 v7, v9, v11, vcc
; VI-NEXT:    v_add_f32_e32 v9, s4, v32
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v5
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_and_b32 s5, s25, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_add_f32_e32 v9, s5, v32
; VI-NEXT:    v_mov_b32_e32 v5, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v11, v13, vcc
; VI-NEXT:    v_bfe_u32 v11, v9, 16, 1
; VI-NEXT:    v_add_u32_e32 v11, vcc, v11, v9
; VI-NEXT:    v_add_u32_e32 v11, vcc, 0x7fff, v11
; VI-NEXT:    s_lshl_b32 s4, s25, 16
; VI-NEXT:    v_or_b32_e32 v13, 0x400000, v9
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v9, v9
; VI-NEXT:    v_cndmask_b32_e32 v9, v11, v13, vcc
; VI-NEXT:    v_add_f32_e32 v11, s4, v32
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v7
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_and_b32 s5, s27, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_add_f32_e32 v11, s5, v32
; VI-NEXT:    v_mov_b32_e32 v7, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v13, v15, vcc
; VI-NEXT:    v_bfe_u32 v13, v11, 16, 1
; VI-NEXT:    v_add_u32_e32 v13, vcc, v13, v11
; VI-NEXT:    v_add_u32_e32 v13, vcc, 0x7fff, v13
; VI-NEXT:    s_lshl_b32 s4, s27, 16
; VI-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v11, v11
; VI-NEXT:    v_cndmask_b32_e32 v11, v13, v15, vcc
; VI-NEXT:    v_add_f32_e32 v13, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v9
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_and_b32 s5, s29, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v9, v18
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_add_f32_e32 v13, s5, v32
; VI-NEXT:    v_cndmask_b32_e32 v18, v15, v18, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v11
; VI-NEXT:    v_bfe_u32 v15, v13, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    v_add_u32_e32 v15, vcc, v15, v13
; VI-NEXT:    v_add_u32_e32 v15, vcc, 0x7fff, v15
; VI-NEXT:    s_lshl_b32 s4, s29, 16
; VI-NEXT:    v_mov_b32_e32 v11, v18
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v13
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v13, v13
; VI-NEXT:    v_cndmask_b32_e32 v13, v15, v18, vcc
; VI-NEXT:    v_add_f32_e32 v15, s4, v32
; VI-NEXT:    v_bfe_u32 v18, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v15
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v18, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v13
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    s_and_b32 s5, s31, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v15, s5, v32
; VI-NEXT:    v_mov_b32_e32 v13, v18
; VI-NEXT:    v_bfe_u32 v18, v15, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v15
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    s_lshl_b32 s4, s31, 16
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v15
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v15, v15
; VI-NEXT:    v_cndmask_b32_e32 v15, v18, v19, vcc
; VI-NEXT:    v_add_f32_e32 v18, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v18
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v15
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[18:19]
; VI-NEXT:    s_and_b32 s5, s50, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v15, v18
; VI-NEXT:    v_add_f32_e32 v18, s5, v32
; VI-NEXT:    v_bfe_u32 v19, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v18
; VI-NEXT:    s_lshl_b32 s4, s50, 16
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_add_f32_e32 v18, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v19, v19, v20, vcc
; VI-NEXT:    v_bfe_u32 v20, v18, 16, 1
; VI-NEXT:    v_add_u32_e32 v20, vcc, v20, v18
; VI-NEXT:    s_and_b32 s4, s46, 0xffff0000
; VI-NEXT:    v_add_u32_e32 v20, vcc, 0x7fff, v20
; VI-NEXT:    v_lshrrev_b64 v[28:29], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v18
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v18, v20, v21, vcc
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s46, 16
; VI-NEXT:    v_lshrrev_b64 v[30:31], 16, v[18:19]
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s44, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[26:27], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s42, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[24:25], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s40, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[22:23], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s38, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[20:21], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v18, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v18, vcc
; VI-NEXT:    v_bfe_u32 v18, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v18, vcc, v18, v16
; VI-NEXT:    v_add_u32_e32 v18, vcc, 0x7fff, v18
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v18, v19, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    s_and_b32 s4, s36, 0xffff0000
; VI-NEXT:    v_lshrrev_b64 v[18:19], 16, v[16:17]
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_bfe_u32 v17, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v17, vcc, v17, v16
; VI-NEXT:    v_add_u32_e32 v17, vcc, 0x7fff, v17
; VI-NEXT:    s_lshl_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v19, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_add_f32_e32 v16, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v17, v17, v19, vcc
; VI-NEXT:    v_bfe_u32 v19, v16, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v16
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v16
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; VI-NEXT:    v_cndmask_b32_e32 v16, v19, v21, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; VI-NEXT:    v_lshrrev_b64 v[16:17], 16, v[16:17]
; VI-NEXT:    s_and_b32 s4, s37, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v17, s4, v32
; VI-NEXT:    v_bfe_u32 v19, v17, 16, 1
; VI-NEXT:    v_add_u32_e32 v19, vcc, v19, v17
; VI-NEXT:    v_add_u32_e32 v19, vcc, 0x7fff, v19
; VI-NEXT:    v_or_b32_e32 v21, 0x400000, v17
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; VI-NEXT:    s_lshl_b32 s4, s37, 16
; VI-NEXT:    v_cndmask_b32_e32 v17, v19, v21, vcc
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    s_and_b32 s4, s39, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    v_add_f32_e32 v19, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v21, v23, vcc
; VI-NEXT:    v_bfe_u32 v21, v19, 16, 1
; VI-NEXT:    v_add_u32_e32 v21, vcc, v21, v19
; VI-NEXT:    v_add_u32_e32 v21, vcc, 0x7fff, v21
; VI-NEXT:    v_or_b32_e32 v23, 0x400000, v19
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; VI-NEXT:    s_lshl_b32 s4, s39, 16
; VI-NEXT:    v_cndmask_b32_e32 v19, v21, v23, vcc
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v17
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    s_and_b32 s4, s41, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    v_add_f32_e32 v21, s4, v32
; VI-NEXT:    v_mov_b32_e32 v17, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v23, v25, vcc
; VI-NEXT:    v_bfe_u32 v23, v21, 16, 1
; VI-NEXT:    v_add_u32_e32 v23, vcc, v23, v21
; VI-NEXT:    v_add_u32_e32 v23, vcc, 0x7fff, v23
; VI-NEXT:    v_or_b32_e32 v25, 0x400000, v21
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v21, v21
; VI-NEXT:    s_lshl_b32 s4, s41, 16
; VI-NEXT:    v_cndmask_b32_e32 v21, v23, v25, vcc
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v19
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    s_and_b32 s4, s43, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    v_add_f32_e32 v23, s4, v32
; VI-NEXT:    v_mov_b32_e32 v19, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v25, v27, vcc
; VI-NEXT:    v_bfe_u32 v25, v23, 16, 1
; VI-NEXT:    v_add_u32_e32 v25, vcc, v25, v23
; VI-NEXT:    v_add_u32_e32 v25, vcc, 0x7fff, v25
; VI-NEXT:    v_or_b32_e32 v27, 0x400000, v23
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v23, v23
; VI-NEXT:    s_lshl_b32 s4, s43, 16
; VI-NEXT:    v_cndmask_b32_e32 v23, v25, v27, vcc
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v21
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    s_and_b32 s4, s45, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    v_add_f32_e32 v25, s4, v32
; VI-NEXT:    v_mov_b32_e32 v21, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v27, v29, vcc
; VI-NEXT:    v_bfe_u32 v27, v25, 16, 1
; VI-NEXT:    v_add_u32_e32 v27, vcc, v27, v25
; VI-NEXT:    v_add_u32_e32 v27, vcc, 0x7fff, v27
; VI-NEXT:    v_or_b32_e32 v29, 0x400000, v25
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v25, v25
; VI-NEXT:    s_lshl_b32 s4, s45, 16
; VI-NEXT:    v_cndmask_b32_e32 v25, v27, v29, vcc
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v23
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    v_add_f32_e32 v27, s4, v32
; VI-NEXT:    v_mov_b32_e32 v23, v33
; VI-NEXT:    v_cndmask_b32_e32 v33, v29, v31, vcc
; VI-NEXT:    v_bfe_u32 v29, v27, 16, 1
; VI-NEXT:    v_add_u32_e32 v29, vcc, v29, v27
; VI-NEXT:    v_add_u32_e32 v29, vcc, 0x7fff, v29
; VI-NEXT:    v_or_b32_e32 v31, 0x400000, v27
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v27, v27
; VI-NEXT:    s_lshl_b32 s4, s47, 16
; VI-NEXT:    v_cndmask_b32_e32 v27, v29, v31, vcc
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    s_and_b32 s4, s49, 0xffff0000
; VI-NEXT:    v_mov_b32_e32 v25, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    v_add_f32_e32 v29, s4, v32
; VI-NEXT:    v_cndmask_b32_e32 v33, v31, v33, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v27
; VI-NEXT:    v_bfe_u32 v31, v29, 16, 1
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    v_add_u32_e32 v31, vcc, v31, v29
; VI-NEXT:    v_add_u32_e32 v31, vcc, 0x7fff, v31
; VI-NEXT:    v_mov_b32_e32 v27, v33
; VI-NEXT:    v_or_b32_e32 v33, 0x400000, v29
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v29, v29
; VI-NEXT:    s_lshl_b32 s4, s49, 16
; VI-NEXT:    v_cndmask_b32_e32 v29, v31, v33, vcc
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v34, 16, v29
; VI-NEXT:    v_lshrrev_b64 v[33:34], 16, v[33:34]
; VI-NEXT:    s_and_b32 s4, s51, 0xffff0000
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_mov_b32_e32 v29, v33
; VI-NEXT:    v_bfe_u32 v33, v31, 16, 1
; VI-NEXT:    v_add_u32_e32 v33, vcc, v33, v31
; VI-NEXT:    v_add_u32_e32 v33, vcc, 0x7fff, v33
; VI-NEXT:    s_lshl_b32 s4, s51, 16
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_add_f32_e32 v31, s4, v32
; VI-NEXT:    v_bfe_u32 v32, v31, 16, 1
; VI-NEXT:    v_cndmask_b32_e32 v33, v33, v34, vcc
; VI-NEXT:    v_add_u32_e32 v32, vcc, v32, v31
; VI-NEXT:    v_add_u32_e32 v32, vcc, 0x7fff, v32
; VI-NEXT:    v_or_b32_e32 v34, 0x400000, v31
; VI-NEXT:    v_cmp_u_f32_e32 vcc, v31, v31
; VI-NEXT:    v_cndmask_b32_e32 v31, v32, v34, vcc
; VI-NEXT:    v_lshrrev_b32_e32 v32, 16, v33
; VI-NEXT:    v_lshrrev_b64 v[31:32], 16, v[31:32]
; VI-NEXT:    s_branch .LBB105_5
; VI-NEXT:  .LBB105_3:
; VI-NEXT:    s_branch .LBB105_2
; VI-NEXT:  .LBB105_4:
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:  .LBB105_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v35, 9
; VI-NEXT:    v_readlane_b32 s50, v35, 8
; VI-NEXT:    v_readlane_b32 s49, v35, 7
; VI-NEXT:    v_readlane_b32 s48, v35, 6
; VI-NEXT:    v_readlane_b32 s39, v35, 5
; VI-NEXT:    v_readlane_b32 s38, v35, 4
; VI-NEXT:    v_readlane_b32 s37, v35, 3
; VI-NEXT:    v_readlane_b32 s36, v35, 2
; VI-NEXT:    v_readlane_b32 s31, v35, 1
; VI-NEXT:    v_readlane_b32 s30, v35, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v35, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64bf16_to_v64i16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v36, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v36, s30, 0
; GFX9-NEXT:    v_writelane_b32 v36, s31, 1
; GFX9-NEXT:    v_writelane_b32 v36, s36, 2
; GFX9-NEXT:    v_writelane_b32 v36, s37, 3
; GFX9-NEXT:    v_writelane_b32 v36, s38, 4
; GFX9-NEXT:    v_writelane_b32 v36, s39, 5
; GFX9-NEXT:    v_writelane_b32 v36, s48, 6
; GFX9-NEXT:    v_writelane_b32 v36, s49, 7
; GFX9-NEXT:    v_writelane_b32 v36, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v36, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB105_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB105_4
; GFX9-NEXT:  .LBB105_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x40c00000
; GFX9-NEXT:    s_and_b32 s4, s50, 0xffff0000
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    s_and_b32 s5, s30, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    v_add_f32_e32 v0, s5, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v1, v2, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s5, s30, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s5, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    s_and_b32 s5, s31, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    v_add_f32_e32 v2, s5, v16
; GFX9-NEXT:    v_bfe_u32 v3, v2, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v3, v3, v2
; GFX9-NEXT:    v_add_u32_e32 v3, 0x7fff, v3
; GFX9-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v2, v2
; GFX9-NEXT:    s_lshl_b32 s5, s31, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v2, v3, v4, vcc
; GFX9-NEXT:    v_add_f32_e32 v3, s5, v16
; GFX9-NEXT:    v_bfe_u32 v4, v3, 16, 1
; GFX9-NEXT:    s_and_b32 s4, s29, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v4, v4, v3
; GFX9-NEXT:    v_mov_b32_e32 v32, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_add_u32_e32 v4, 0x7fff, v4
; GFX9-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v3, v3
; GFX9-NEXT:    v_and_or_b32 v14, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_and_or_b32 v15, v2, v32, v3
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s29, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s28, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v13, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s28, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s27, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v12, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s27, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s26, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v11, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s26, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s25, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v10, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s25, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s24, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v9, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s24, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s23, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v8, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s23, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s22, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v7, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s22, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s21, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v6, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s21, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s20, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v5, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s20, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s19, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v4, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s19, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
; GFX9-NEXT:    s_and_b32 s4, s18, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v3, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v2, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s18, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v2, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v2, v2, v1
; GFX9-NEXT:    v_add_u32_e32 v2, 0x7fff, v2
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v18, vcc
; GFX9-NEXT:    s_and_b32 s4, s17, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v2, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v1, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v1, v1, v0
; GFX9-NEXT:    v_add_u32_e32 v1, 0x7fff, v1
; GFX9-NEXT:    v_or_b32_e32 v18, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s17, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v1, v18, vcc
; GFX9-NEXT:    v_add_f32_e32 v1, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v1, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v1
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v1
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v1, v1
; GFX9-NEXT:    v_cndmask_b32_e32 v1, v18, v19, vcc
; GFX9-NEXT:    s_and_b32 s4, s16, 0xffff0000
; GFX9-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; GFX9-NEXT:    v_and_or_b32 v1, v0, v32, v1
; GFX9-NEXT:    v_add_f32_e32 v0, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v0, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v0
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v0
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v0, v0
; GFX9-NEXT:    s_lshl_b32 s4, s16, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v0, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    s_lshl_b32 s4, s50, 16
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    v_and_or_b32 v0, v0, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    s_and_b32 s4, s51, 0xffff0000
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_add_f32_e32 v19, s4, v16
; GFX9-NEXT:    v_bfe_u32 v20, v19, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v20, v20, v19
; GFX9-NEXT:    v_add_u32_e32 v20, 0x7fff, v20
; GFX9-NEXT:    v_or_b32_e32 v21, 0x400000, v19
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v19, v19
; GFX9-NEXT:    s_lshl_b32 s4, s51, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v19, v20, v21, vcc
; GFX9-NEXT:    v_add_f32_e32 v20, s4, v16
; GFX9-NEXT:    v_bfe_u32 v21, v20, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v21, v21, v20
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s49, 0xffff0000
; GFX9-NEXT:    v_add_u32_e32 v21, 0x7fff, v21
; GFX9-NEXT:    v_or_b32_e32 v22, 0x400000, v20
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v20, v20
; GFX9-NEXT:    v_and_or_b32 v30, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v20, v21, v22, vcc
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_and_or_b32 v31, v19, v32, v20
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s49, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s48, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v29, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s48, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s47, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v28, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s47, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s46, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v27, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s46, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s45, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v26, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s45, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s44, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v25, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s44, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s43, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v24, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s43, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s42, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v23, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s42, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s41, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v22, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s41, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s40, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v21, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s40, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v20, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v20, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s39, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v20, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v19, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s39, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v19, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v19, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v19, v19, v18
; GFX9-NEXT:    v_add_u32_e32 v19, 0x7fff, v19
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v19, v33, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s38, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v19, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v18, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v18, v18, v17
; GFX9-NEXT:    v_add_u32_e32 v18, 0x7fff, v18
; GFX9-NEXT:    v_or_b32_e32 v33, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s38, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v18, v33, vcc
; GFX9-NEXT:    v_add_f32_e32 v18, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v18, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v18
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v18
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v18, v18
; GFX9-NEXT:    v_cndmask_b32_e32 v18, v33, v34, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; GFX9-NEXT:    s_and_b32 s4, s37, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v18, v17, v32, v18
; GFX9-NEXT:    v_add_f32_e32 v17, s4, v16
; GFX9-NEXT:    v_bfe_u32 v33, v17, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v33, v33, v17
; GFX9-NEXT:    v_add_u32_e32 v33, 0x7fff, v33
; GFX9-NEXT:    v_or_b32_e32 v34, 0x400000, v17
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v17, v17
; GFX9-NEXT:    s_lshl_b32 s4, s37, 16
; GFX9-NEXT:    v_cndmask_b32_e32 v17, v33, v34, vcc
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v33, 16, v33
; GFX9-NEXT:    s_and_b32 s4, s36, 0xffff0000
; GFX9-NEXT:    v_and_or_b32 v17, v17, v32, v33
; GFX9-NEXT:    v_add_f32_e32 v33, s4, v16
; GFX9-NEXT:    v_bfe_u32 v34, v33, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v33
; GFX9-NEXT:    s_lshl_b32 s4, s36, 16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v33
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v33, v33
; GFX9-NEXT:    v_add_f32_e32 v16, s4, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v33, v34, v35, vcc
; GFX9-NEXT:    v_bfe_u32 v34, v16, 16, 1
; GFX9-NEXT:    v_add_u32_e32 v34, v34, v16
; GFX9-NEXT:    v_add_u32_e32 v34, 0x7fff, v34
; GFX9-NEXT:    v_or_b32_e32 v35, 0x400000, v16
; GFX9-NEXT:    v_cmp_u_f32_e32 vcc, v16, v16
; GFX9-NEXT:    v_cndmask_b32_e32 v16, v34, v35, vcc
; GFX9-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; GFX9-NEXT:    v_and_or_b32 v16, v33, v32, v16
; GFX9-NEXT:    s_branch .LBB105_5
; GFX9-NEXT:  .LBB105_3:
; GFX9-NEXT:    s_branch .LBB105_2
; GFX9-NEXT:  .LBB105_4:
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:  .LBB105_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v36, 9
; GFX9-NEXT:    v_readlane_b32 s50, v36, 8
; GFX9-NEXT:    v_readlane_b32 s49, v36, 7
; GFX9-NEXT:    v_readlane_b32 s48, v36, 6
; GFX9-NEXT:    v_readlane_b32 s39, v36, 5
; GFX9-NEXT:    v_readlane_b32 s38, v36, 4
; GFX9-NEXT:    v_readlane_b32 s37, v36, 3
; GFX9-NEXT:    v_readlane_b32 s36, v36, 2
; GFX9-NEXT:    v_readlane_b32 s31, v36, 1
; GFX9-NEXT:    v_readlane_b32 s30, v36, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v36, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-TRUE16-LABEL: bitcast_v64bf16_to_v64i16_scalar:
; GFX11-TRUE16:       ; %bb.0:
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-TRUE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s12, s0
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-TRUE16-NEXT:    s_mov_b32 s37, s29
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-TRUE16-NEXT:    s_mov_b32 s36, s28
; GFX11-TRUE16-NEXT:    s_mov_b32 s15, s3
; GFX11-TRUE16-NEXT:    s_mov_b32 s14, s2
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-TRUE16-NEXT:    s_mov_b32 s13, s1
; GFX11-TRUE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-TRUE16-NEXT:    s_mov_b32 s0, 0
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-TRUE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-TRUE16-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-TRUE16-NEXT:    s_cbranch_scc0 .LBB105_3
; GFX11-TRUE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-TRUE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-TRUE16-NEXT:    s_cbranch_vccnz .LBB105_4
; GFX11-TRUE16-NEXT:  .LBB105_2: ; %cmp.true
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-TRUE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-TRUE16-NEXT:    s_and_b32 s1, s25, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v0, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s37, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v2, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s38, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v2
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v2
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v16, v3, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v10, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v3, 16, 1
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v3
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v32, v5, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v16.l, v32.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v17, v6, v9, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v33, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v3, 0x400000, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v6
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v18, v1, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v7
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v34, v0, v3, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v19, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v35, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v20, v1, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v36, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v8, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v21, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v37, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v22, v1, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v38, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v8, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v23, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v39, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v24, v1, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v48, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v8, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v25, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v49, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v26, v1, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v50, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v8, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v27, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v51, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v28, v1, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s50, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v52, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v8, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v29, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v53, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, v0, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v30, v1, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v0
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s12, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v54, v0, v5, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v1, v8, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v31, v2, v7, vcc_lo
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, v1, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s13, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v5, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v55, v0, v2, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v0, v3, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v4, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v0, v3
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s14, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, v2, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v0, v1, v5 :: v_dual_add_nc_u32 v1, 0x7fff, v6
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v3
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v4
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v7
; GFX11-TRUE16-NEXT:    v_bfe_u32 v3, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v64, v1, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s14, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, v3, v9
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v0.l, v64.h
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v1, v2, v8 :: v_dual_add_nc_u32 v2, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v3, 0x7fff, v3
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s15, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v65, v2, v4, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v2, v5, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v7, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v2, v5
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v2, v3, v7 :: v_dual_add_nc_u32 v3, 0x7fff, v8
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v10
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s16, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v66, v3, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-TRUE16-NEXT:    v_bfe_u32 v5, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v6, 0x400000, v10
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v1.l, v65.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v2.l, v66.h
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v3, v4, v9 :: v_dual_add_nc_u32 v4, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, v5, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s17, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v67, v4, v6, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v8, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v9, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v4, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, v6, v8
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v4, v5, v9 :: v_dual_add_nc_u32 v5, 0x7fff, v10
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v7
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v6
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s18, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v7, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v68, v5, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v8, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v3.l, v67.h
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, v7, v13
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v4.l, v68.h
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v5, v6, v11 :: v_dual_add_nc_u32 v6, 0x7fff, v9
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v9, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v7
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s19, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v69, v6, v8, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v6, v9, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v11, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v6, v9
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, v8, v10
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v10
; GFX11-TRUE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v6, v7, v11 :: v_dual_add_nc_u32 v7, 0x7fff, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v9
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v8
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v14
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s20, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v70, v7, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT:    v_bfe_u32 v9, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v10, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v5.l, v69.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v6.l, v70.h
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v7, v8, v13 :: v_dual_add_nc_u32 v8, 0x7fff, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, v9, v15
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s21, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v71, v8, v10, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v12, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v13, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v80, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v8, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, v10, v12
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v81, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v8, v9, v13 :: v_dual_add_nc_u32 v9, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v80, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v11
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s22, 16
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, v13, v80
; GFX11-TRUE16-NEXT:    v_bfe_u32 v11, v81, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v82, v9, v14, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v80
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v7.l, v71.h
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v11, v81
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v8.l, v82.h
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v9, v10, v15 :: v_dual_add_nc_u32 v10, 0x7fff, v13
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v80, v80
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s23, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v80, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v12, 0x400000, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v15, v14, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v84, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v83, v10, v13
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v81, v84, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v10, v11, v12, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, v15, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v83
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v15, 0x400000, v13
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v11, 0x7fff, v11
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v81, v81, v84
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s24, 16
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v86, v12, v15, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-TRUE16-NEXT:    v_bfe_u32 v12, v83, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v81
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v14, 0x400000, v84
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v11, v11, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, v12, v83
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v81, 0x40c00000, s1
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s25, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v83
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v84, v13, v14, vcc_lo
; GFX11-TRUE16-NEXT:    v_bfe_u32 v13, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
; GFX11-TRUE16-NEXT:    v_bfe_u32 v14, v81, 16, 1
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v87, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, v13, v15
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, 0x400000, v81
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v14, v81
; GFX11-TRUE16-NEXT:    v_bfe_u32 v83, v87, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v12, v12, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v13, 0x7fff, v13
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v85, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v83, v83, v87
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v11.l, v84.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v85, v13, v85, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v81, v81
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s26, 16
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v10.l, v86.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v12.l, v85.h
; GFX11-TRUE16-NEXT:    v_dual_cndmask_b32 v13, v14, v96 :: v_dual_add_nc_u32 v14, 0x7fff, v83
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v83, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    s_lshl_b32 s0, s27, 16
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v96, 0x400000, v87
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v97, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-TRUE16-NEXT:    v_bfe_u32 v98, v83, 16, 1
; GFX11-TRUE16-NEXT:    s_and_b32 s0, s27, 0xffff0000
; GFX11-TRUE16-NEXT:    v_bfe_u32 v81, v15, 16, 1
; GFX11-TRUE16-NEXT:    v_bfe_u32 v100, v97, 16, 1
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v87, v14, v96, vcc_lo
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, v98, v83
; GFX11-TRUE16-NEXT:    v_add_f32_e64 v99, 0x40c00000, s0
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v101, 0x400000, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v98, v100, v97
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v83, v83
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v81, v81, v15
; GFX11-TRUE16-NEXT:    v_bfe_u32 v96, v99, 16, 1
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v98, 0x7fff, v98
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v102, 0x400000, v97
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v83, v14, v101, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v97, v97
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v81, 0x7fff, v81
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v100, 0x400000, v15
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v96, v96, v99
; GFX11-TRUE16-NEXT:    v_or_b32_e32 v101, 0x400000, v99
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v97, v98, v102, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v13.l, v87.h
; GFX11-TRUE16-NEXT:    v_add_nc_u32_e32 v96, 0x7fff, v96
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v9.l, v80.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v31.l, v55.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v14, v81, v100, vcc_lo
; GFX11-TRUE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v99, v99
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v14.l, v83.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v30.l, v54.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v29.l, v53.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v28.l, v52.h
; GFX11-TRUE16-NEXT:    v_cndmask_b32_e32 v15, v96, v101, vcc_lo
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v15.l, v97.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v27.l, v51.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v26.l, v50.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v25.l, v49.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v24.l, v48.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v23.l, v39.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v22.l, v38.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v21.l, v37.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v20.l, v36.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v19.l, v35.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v18.l, v34.h
; GFX11-TRUE16-NEXT:    v_mov_b16_e32 v17.l, v33.h
; GFX11-TRUE16-NEXT:    s_branch .LBB105_5
; GFX11-TRUE16-NEXT:  .LBB105_3:
; GFX11-TRUE16-NEXT:    s_branch .LBB105_2
; GFX11-TRUE16-NEXT:  .LBB105_4:
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-TRUE16-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-TRUE16-NEXT:  .LBB105_5: ; %end
; GFX11-TRUE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-TRUE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-TRUE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-TRUE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-TRUE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-TRUE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-TRUE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-TRUE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-TRUE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-TRUE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-TRUE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-TRUE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-TRUE16-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v64bf16_to_v64i16_scalar:
; GFX11-FAKE16:       ; %bb.0:
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s4, -1
; GFX11-FAKE16-NEXT:    scratch_store_b32 off, v40, s32 ; 4-byte Folded Spill
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s36, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s12, s0
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s37, 1
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s38, 2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-FAKE16-NEXT:    s_mov_b32 s37, s29
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s39, 3
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-FAKE16-NEXT:    s_mov_b32 s36, s28
; GFX11-FAKE16-NEXT:    s_mov_b32 s15, s3
; GFX11-FAKE16-NEXT:    s_mov_b32 s14, s2
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s48, 4
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-FAKE16-NEXT:    s_mov_b32 s13, s1
; GFX11-FAKE16-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-FAKE16-NEXT:    s_mov_b32 s0, 0
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s49, 5
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s50, 6
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-FAKE16-NEXT:    v_writelane_b32 v40, s51, 7
; GFX11-FAKE16-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-FAKE16-NEXT:    s_cbranch_scc0 .LBB105_3
; GFX11-FAKE16-NEXT:  ; %bb.1: ; %Flow
; GFX11-FAKE16-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-FAKE16-NEXT:    s_cbranch_vccnz .LBB105_4
; GFX11-FAKE16-NEXT:  .LBB105_2: ; %cmp.true
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s36, 16
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s36, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s37, 0xffff0000
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s37, 16
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v2, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_and_b32 s2, s38, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v2, v0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v3
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s38, 16
; GFX11-FAKE16-NEXT:    s_lshl_b32 s1, s24, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v2
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v16, v2, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v1, 0x400000, v3
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, v10, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v17, v4, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_2)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v4, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v32, v0, v1, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v2
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s39, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s39, 16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v18, v1, v2 :: v_dual_add_nc_u32 v1, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v4, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v33, v1, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s40, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s40, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v19, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v34, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s41, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v20, v2, v3 :: v_dual_add_nc_u32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s41, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v35, v1, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s42, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s42, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v36, v2, v3 :: v_dual_add_nc_u32 v3, v5, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v36, 16, v36
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v21, v0, v2 :: v_dual_add_nc_u32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s43, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v22, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s43, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v37, v1, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s44, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s44, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v23, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v38, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s45, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v24, v2, v3 :: v_dual_add_nc_u32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s45, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v39, v1, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s46, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s46, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v25, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v48, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s47, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v49, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s47, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v49, 16, v49
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v26, v1, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s48, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s48, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v27, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v50, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s49, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v28, v2, v3 :: v_dual_add_nc_u32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s49, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v51, v1, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s50, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s50, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v29, v2, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v52, v0, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v0, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s51, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v0, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v30, v2, v3 :: v_dual_add_nc_u32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s51, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v3, v5, v0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v53, v1, v2, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v1, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s12, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v2, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v1, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v0, v0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s12, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v0, 0x7fff, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v4, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v54, v2, v3 :: v_dual_add_nc_u32 v3, v5, v1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v2, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v4, v6
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v54, 16, v54
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v31, v0, v2 :: v_dual_add_nc_u32 v0, 0x7fff, v3
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v2, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v1, v1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s13, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v2, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v0, v0, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v3, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s13, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, v5, v2
; GFX11-FAKE16-NEXT:    v_bfe_u32 v5, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v55, v1, v3, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v3, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s14, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v1, 0x7fff, v4
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v2
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v5, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v3, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v8, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v2, v2
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s14, 16
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v2, v1, v4 :: v_dual_add_nc_u32 v1, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v4, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, v6, v3
; GFX11-FAKE16-NEXT:    v_bfe_u32 v6, v8, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v3
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v5
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, v6, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v1, v1, v4, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v3, v3
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s15, 0xffff0000
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v3, v5, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v6
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v8
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v8, v8
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s15, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v9, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v64, v5, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s16, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v9, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s16, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v4, 0x7fff, v8
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v65, v6, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v9, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v11
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v65
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v66, v4, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v4, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s17, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v4, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v10, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v5, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v67, v6, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v6, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s17, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, v9, v4
; GFX11-FAKE16-NEXT:    v_bfe_u32 v8, v10, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v67, 16, v67
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v68, v5, v6, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v5, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s18, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v7
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v4
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v8, v10
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v5, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v4, v4
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s18, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v4, v6, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v6, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v7, 0x400000, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, v9, v5
; GFX11-FAKE16-NEXT:    v_bfe_u32 v9, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; GFX11-FAKE16-NEXT:    s_delay_alu instid0(VALU_DEP_3)
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v9, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v69, v6, v7, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v6, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v8
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v5
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v5, v5
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s19, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v6, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s19, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v5, v7, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v7, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v8, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v10, v6
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v14, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v70, v7, v8, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v7, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s20, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v9, 0x400000, v6
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v11, v7, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v6, v6
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s20, 16
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v7
; GFX11-FAKE16-NEXT:    v_and_or_b32 v5, 0xffff0000, v69, v5
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v69, 16, v2
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v6, v8, v9, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v11, v7
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s21, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v80, v8, v14, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v7, v7
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v12, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v8, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v10, 0x400000, v13
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s21, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v7, v9, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v71, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v12, 16, 1
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v8, v8, v10 :: v_dual_add_nc_u32 v9, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s22, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v10, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v11, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s22, 16
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v81, v9, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, v71, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, v10, v14
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v15, 0x400000, v12
; GFX11-FAKE16-NEXT:    v_bfe_u32 v71, v11, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v9, 0x7fff, v9
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v12, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v10, 0x7fff, v10
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v71, v11
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v9, v9, v15, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s23, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v83, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v15, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v10, v10, v82, vcc_lo
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v71, 0x400000, v11
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s23, 16
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v82, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_bfe_u32 v83, v15, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v84, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v71, v12, v71, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s24, 0xffff0000
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, v83, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v13, v84, 16, 1
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v83, 0x400000, v15
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v11, v14, v82, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v14, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v82, 0x40c00000, s1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v12, 0x7fff, v12
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v13, v13, v84
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v15, v15
; GFX11-FAKE16-NEXT:    v_bfe_u32 v85, v14, 16, 1
; GFX11-FAKE16-NEXT:    v_bfe_u32 v86, v82, 16, 1
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s25, 0xffff0000
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v96, 0x400000, v82
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v83, v12, v83 :: v_dual_add_nc_u32 v12, 0x7fff, v13
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v13, 0x400000, v84
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v85, v14
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v84, v84
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v85, v86, v82
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v84, 0x400000, v14
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; GFX11-FAKE16-NEXT:    v_dual_cndmask_b32 v12, v12, v13 :: v_dual_add_nc_u32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v14, v14
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v85, 0x7fff, v85
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v13, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s25, 16
; GFX11-FAKE16-NEXT:    v_and_or_b32 v6, 0xffff0000, v70, v6
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v84, v15, v84, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v82, v82
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v87, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s26, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v86, v13, 16, 1
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v70, 16, v0
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v82, v85, v96, vcc_lo
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v85, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s26, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v15, v87, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v96, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_lshl_b32 s0, s27, 16
; GFX11-FAKE16-NEXT:    v_bfe_u32 v97, v85, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v98, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    s_and_b32 s0, s27, 0xffff0000
; GFX11-FAKE16-NEXT:    v_bfe_u32 v99, v96, 16, 1
; GFX11-FAKE16-NEXT:    v_add_f32_e64 v100, 0x40c00000, s0
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v113, 0x400000, v96
; GFX11-FAKE16-NEXT:    v_bfe_u32 v101, v98, 16, 1
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v96, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, v99, v96
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, v97, v85
; GFX11-FAKE16-NEXT:    v_bfe_u32 v103, v100, 16, 1
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v101, v101, v98
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v114, 0x400000, v98
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, 0x7fff, v99
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v97, 0x7fff, v97
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v112, 0x400000, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v101, 0x7fff, v101
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v103, v103, v100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v96, v99, v113, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v98, v98
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, v15, v87
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, v86, v13
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v99, 0x7fff, v103
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v103, 0x400000, v100
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v98, v101, v114, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v85, v85
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v15, 0x7fff, v15
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v102, 0x400000, v87
; GFX11-FAKE16-NEXT:    v_add_nc_u32_e32 v14, 0x7fff, v14
; GFX11-FAKE16-NEXT:    v_or_b32_e32 v86, 0x400000, v13
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v85, v97, v112, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v100, v100
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v96
; GFX11-FAKE16-NEXT:    v_and_or_b32 v2, 0xffff0000, v64, v65
; GFX11-FAKE16-NEXT:    v_and_or_b32 v0, 0xffff0000, v55, v69
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v55, 16, v30
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v97, v99, v103, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v87, v87
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v64, 16, v29
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v65, 16, v28
; GFX11-FAKE16-NEXT:    v_and_or_b32 v4, 0xffff0000, v68, v4
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v68, 16, v3
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v87, v15, v102, vcc_lo
; GFX11-FAKE16-NEXT:    v_cmp_u_f32_e32 vcc_lo, v13, v13
; GFX11-FAKE16-NEXT:    v_and_or_b32 v3, 0xffff0000, v66, v67
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v66, 16, v27
; GFX11-FAKE16-NEXT:    v_and_or_b32 v29, 0xffff0000, v52, v55
; GFX11-FAKE16-NEXT:    v_and_or_b32 v28, 0xffff0000, v51, v64
; GFX11-FAKE16-NEXT:    v_cndmask_b32_e32 v13, v14, v86, vcc_lo
; GFX11-FAKE16-NEXT:    v_and_or_b32 v14, 0xffff0000, v85, v96
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v85, 16, v87
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v87, 16, v11
; GFX11-FAKE16-NEXT:    v_and_or_b32 v27, 0xffff0000, v50, v65
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v50, 16, v25
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v51, 16, v24
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v52, 16, v23
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v98, 16, v98
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v82, 16, v82
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v86, 16, v12
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v96, 16, v10
; GFX11-FAKE16-NEXT:    v_and_or_b32 v10, 0xffff0000, v71, v87
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v71, 16, v81
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; GFX11-FAKE16-NEXT:    v_and_or_b32 v30, 0xffff0000, v53, v54
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v53, 16, v22
; GFX11-FAKE16-NEXT:    v_and_or_b32 v25, 0xffff0000, v48, v49
; GFX11-FAKE16-NEXT:    v_and_or_b32 v24, 0xffff0000, v39, v50
; GFX11-FAKE16-NEXT:    v_and_or_b32 v23, 0xffff0000, v38, v51
; GFX11-FAKE16-NEXT:    v_and_or_b32 v22, 0xffff0000, v37, v52
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v37, 16, v20
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v38, 16, v19
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v39, 16, v18
; GFX11-FAKE16-NEXT:    v_lshrrev_b32_e32 v48, 16, v17
; GFX11-FAKE16-NEXT:    v_and_or_b32 v15, 0xffff0000, v97, v98
; GFX11-FAKE16-NEXT:    v_and_or_b32 v13, 0xffff0000, v13, v85
; GFX11-FAKE16-NEXT:    v_and_or_b32 v12, 0xffff0000, v84, v82
; GFX11-FAKE16-NEXT:    v_and_or_b32 v11, 0xffff0000, v83, v86
; GFX11-FAKE16-NEXT:    v_and_or_b32 v9, 0xffff0000, v9, v96
; GFX11-FAKE16-NEXT:    v_and_or_b32 v8, 0xffff0000, v8, v71
; GFX11-FAKE16-NEXT:    v_and_or_b32 v7, 0xffff0000, v80, v7
; GFX11-FAKE16-NEXT:    v_and_or_b32 v1, 0xffff0000, v1, v68
; GFX11-FAKE16-NEXT:    v_and_or_b32 v31, 0xffff0000, v31, v70
; GFX11-FAKE16-NEXT:    v_and_or_b32 v26, 0xffff0000, v26, v66
; GFX11-FAKE16-NEXT:    v_and_or_b32 v21, 0xffff0000, v21, v53
; GFX11-FAKE16-NEXT:    v_and_or_b32 v20, 0xffff0000, v35, v36
; GFX11-FAKE16-NEXT:    v_and_or_b32 v19, 0xffff0000, v34, v37
; GFX11-FAKE16-NEXT:    v_and_or_b32 v18, 0xffff0000, v33, v38
; GFX11-FAKE16-NEXT:    v_and_or_b32 v17, 0xffff0000, v32, v39
; GFX11-FAKE16-NEXT:    v_and_or_b32 v16, 0xffff0000, v16, v48
; GFX11-FAKE16-NEXT:    s_branch .LBB105_5
; GFX11-FAKE16-NEXT:  .LBB105_3:
; GFX11-FAKE16-NEXT:    s_branch .LBB105_2
; GFX11-FAKE16-NEXT:  .LBB105_4:
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-FAKE16-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-FAKE16-NEXT:  .LBB105_5: ; %end
; GFX11-FAKE16-NEXT:    v_readlane_b32 s51, v40, 7
; GFX11-FAKE16-NEXT:    v_readlane_b32 s50, v40, 6
; GFX11-FAKE16-NEXT:    v_readlane_b32 s49, v40, 5
; GFX11-FAKE16-NEXT:    v_readlane_b32 s48, v40, 4
; GFX11-FAKE16-NEXT:    v_readlane_b32 s39, v40, 3
; GFX11-FAKE16-NEXT:    v_readlane_b32 s38, v40, 2
; GFX11-FAKE16-NEXT:    v_readlane_b32 s37, v40, 1
; GFX11-FAKE16-NEXT:    v_readlane_b32 s36, v40, 0
; GFX11-FAKE16-NEXT:    s_or_saveexec_b32 s0, -1
; GFX11-FAKE16-NEXT:    scratch_load_b32 v40, off, s32 ; 4-byte Folded Reload
; GFX11-FAKE16-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-FAKE16-NEXT:    s_waitcnt vmcnt(0)
; GFX11-FAKE16-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x bfloat> %a, splat (bfloat 0xR40C0)
  %a2 = bitcast <64 x bfloat> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <64 x bfloat> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define <64 x bfloat> @bitcast_v64i16_to_v64bf16(<64 x i16> %a, i32 %b) {
; SI-LABEL: bitcast_v64i16_to_v64bf16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v22
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v20
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v18
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v16
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v13
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v11
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v9
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v7
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v5
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v3
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v1
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v30
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v33
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v36
; SI-NEXT:    v_lshlrev_b32_e32 v37, 16, v37
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v39
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v48
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v50
; SI-NEXT:    v_lshlrev_b32_e32 v51, 16, v51
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; SI-NEXT:    v_lshlrev_b32_e32 v53, 16, v53
; SI-NEXT:    v_lshlrev_b32_e32 v54, 16, v54
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; SI-NEXT:    v_lshlrev_b32_e32 v40, 16, v40
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v41
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v43
; SI-NEXT:    v_lshlrev_b32_e32 v44, 16, v44
; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v45
; SI-NEXT:    v_lshlrev_b32_e32 v46, 16, v46
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v56
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v57
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v58
; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v59
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v60
; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v61
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v0
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    buffer_store_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v63
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v62
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    ; kill: killed $vgpr31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB106_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v17
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v19
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v21
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v23
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v25
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v29
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v30
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v63
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr63
; SI-NEXT:  .LBB106_2: ; %Flow
; SI-NEXT:    s_or_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_xor_b64 exec, exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB106_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_add_i32_e32 v1, vcc, 3, v1
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_mov_b32 s6, 0x30000
; SI-NEXT:    v_add_i32_e32 v2, vcc, 3, v2
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v31, v0
; SI-NEXT:    v_add_i32_e32 v3, vcc, 3, v3
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v1, v32, v1
; SI-NEXT:    v_add_i32_e32 v0, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v4, vcc, 3, v4
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_or_b32_e32 v2, v33, v2
; SI-NEXT:    v_add_i32_e32 v1, vcc, s6, v1
; SI-NEXT:    v_and_b32_e32 v31, 0xffff0000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_add_i32_e32 v5, vcc, 3, v5
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v3, v34, v3
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v1
; SI-NEXT:    v_add_i32_e32 v6, vcc, 3, v6
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_or_b32_e32 v4, v35, v4
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v3
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v2
; SI-NEXT:    v_add_i32_e32 v7, vcc, 3, v7
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v5, v36, v5
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v4
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v3
; SI-NEXT:    v_add_i32_e32 v8, vcc, 3, v8
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_or_b32_e32 v6, v37, v6
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v5
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v4
; SI-NEXT:    v_add_i32_e32 v9, vcc, 3, v9
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v7, v38, v7
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v5
; SI-NEXT:    v_add_i32_e32 v10, vcc, 3, v10
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_or_b32_e32 v8, v39, v8
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v7
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v6
; SI-NEXT:    v_add_i32_e32 v11, vcc, 3, v11
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v9, v48, v9
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v7
; SI-NEXT:    v_add_i32_e32 v12, vcc, 3, v12
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v10, v49, v10
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v9
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v8
; SI-NEXT:    v_add_i32_e32 v13, vcc, 3, v13
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v11, v50, v11
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v9
; SI-NEXT:    v_add_i32_e32 v14, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_or_b32_e32 v12, v51, v12
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v11
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v10
; SI-NEXT:    v_add_i32_e32 v15, vcc, 3, v15
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v13, v52, v13
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v11
; SI-NEXT:    v_add_i32_e32 v16, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_or_b32_e32 v14, v53, v14
; SI-NEXT:    v_add_i32_e32 v13, vcc, s6, v13
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v12
; SI-NEXT:    v_add_i32_e32 v17, vcc, 3, v17
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v15, v54, v15
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v14
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v13
; SI-NEXT:    v_add_i32_e32 v18, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_or_b32_e32 v16, v55, v16
; SI-NEXT:    v_add_i32_e32 v15, vcc, s6, v15
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v14
; SI-NEXT:    v_add_i32_e32 v19, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v17, v40, v17
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v15
; SI-NEXT:    v_add_i32_e32 v20, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v18, v41, v18
; SI-NEXT:    v_add_i32_e32 v17, vcc, s6, v17
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v16
; SI-NEXT:    v_add_i32_e32 v21, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v19, v42, v19
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v18
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v17
; SI-NEXT:    v_add_i32_e32 v22, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v20, v43, v20
; SI-NEXT:    v_add_i32_e32 v19, vcc, s6, v19
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v18
; SI-NEXT:    v_add_i32_e32 v23, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v21, v44, v21
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v20
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v19
; SI-NEXT:    v_add_i32_e32 v24, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v22, v45, v22
; SI-NEXT:    v_add_i32_e32 v21, vcc, s6, v21
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v20
; SI-NEXT:    v_add_i32_e32 v25, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v23, v46, v23
; SI-NEXT:    v_add_i32_e32 v22, vcc, s6, v22
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v21
; SI-NEXT:    v_add_i32_e32 v26, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v24, v47, v24
; SI-NEXT:    v_add_i32_e32 v23, vcc, s6, v23
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v22
; SI-NEXT:    v_add_i32_e32 v27, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v25, v56, v25
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v24
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v23
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_add_i32_e32 v29, vcc, 3, v29
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v26, v57, v26
; SI-NEXT:    v_add_i32_e32 v25, vcc, s6, v25
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v24
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v27, v58, v27
; SI-NEXT:    v_add_i32_e32 v26, vcc, s6, v26
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v25
; SI-NEXT:    v_or_b32_e32 v30, v61, v30
; SI-NEXT:    v_or_b32_e32 v60, v60, v29
; SI-NEXT:    v_or_b32_e32 v59, v59, v28
; SI-NEXT:    v_add_i32_e32 v27, vcc, s6, v27
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v26
; SI-NEXT:    v_add_i32_e32 v63, vcc, 3, v63
; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v30
; SI-NEXT:    v_add_i32_e32 v30, vcc, s6, v60
; SI-NEXT:    v_add_i32_e32 v60, vcc, s6, v59
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v27
; SI-NEXT:    v_and_b32_e32 v63, 0xffff, v63
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v60
; SI-NEXT:    v_or_b32_e32 v62, v62, v63
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v30
; SI-NEXT:    v_add_i32_e32 v28, vcc, 0x30000, v62
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v29
; SI-NEXT:    v_and_b32_e32 v32, 0xffff0000, v1
; SI-NEXT:    v_and_b32_e32 v33, 0xffff0000, v2
; SI-NEXT:    v_and_b32_e32 v34, 0xffff0000, v3
; SI-NEXT:    v_and_b32_e32 v35, 0xffff0000, v4
; SI-NEXT:    v_and_b32_e32 v36, 0xffff0000, v5
; SI-NEXT:    v_and_b32_e32 v37, 0xffff0000, v6
; SI-NEXT:    v_and_b32_e32 v38, 0xffff0000, v7
; SI-NEXT:    v_and_b32_e32 v39, 0xffff0000, v8
; SI-NEXT:    v_and_b32_e32 v48, 0xffff0000, v9
; SI-NEXT:    v_and_b32_e32 v49, 0xffff0000, v10
; SI-NEXT:    v_and_b32_e32 v50, 0xffff0000, v11
; SI-NEXT:    v_and_b32_e32 v51, 0xffff0000, v12
; SI-NEXT:    v_and_b32_e32 v52, 0xffff0000, v13
; SI-NEXT:    v_and_b32_e32 v53, 0xffff0000, v14
; SI-NEXT:    v_and_b32_e32 v54, 0xffff0000, v15
; SI-NEXT:    v_and_b32_e32 v55, 0xffff0000, v16
; SI-NEXT:    v_and_b32_e32 v40, 0xffff0000, v17
; SI-NEXT:    v_and_b32_e32 v41, 0xffff0000, v18
; SI-NEXT:    v_and_b32_e32 v42, 0xffff0000, v19
; SI-NEXT:    v_and_b32_e32 v43, 0xffff0000, v20
; SI-NEXT:    v_and_b32_e32 v44, 0xffff0000, v21
; SI-NEXT:    v_and_b32_e32 v45, 0xffff0000, v22
; SI-NEXT:    v_and_b32_e32 v46, 0xffff0000, v23
; SI-NEXT:    v_and_b32_e32 v47, 0xffff0000, v24
; SI-NEXT:    v_and_b32_e32 v56, 0xffff0000, v25
; SI-NEXT:    v_and_b32_e32 v57, 0xffff0000, v26
; SI-NEXT:    v_and_b32_e32 v58, 0xffff0000, v27
; SI-NEXT:    v_and_b32_e32 v59, 0xffff0000, v60
; SI-NEXT:    v_and_b32_e32 v60, 0xffff0000, v30
; SI-NEXT:    v_and_b32_e32 v61, 0xffff0000, v29
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    v_and_b32_e32 v62, 0xffff0000, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:  .LBB106_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14) expcnt(0)
; SI-NEXT:    v_mul_f32_e32 v0, 1.0, v31
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v1
; SI-NEXT:    v_alignbit_b32 v0, v0, v1, 16
; SI-NEXT:    v_mul_f32_e32 v1, 1.0, v32
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v2
; SI-NEXT:    v_alignbit_b32 v1, v1, v2, 16
; SI-NEXT:    v_mul_f32_e32 v2, 1.0, v33
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v3
; SI-NEXT:    v_alignbit_b32 v2, v2, v3, 16
; SI-NEXT:    v_mul_f32_e32 v3, 1.0, v34
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v4
; SI-NEXT:    v_alignbit_b32 v3, v3, v4, 16
; SI-NEXT:    v_mul_f32_e32 v4, 1.0, v35
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v5
; SI-NEXT:    v_alignbit_b32 v4, v4, v5, 16
; SI-NEXT:    v_mul_f32_e32 v5, 1.0, v36
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v6
; SI-NEXT:    v_alignbit_b32 v5, v5, v6, 16
; SI-NEXT:    v_mul_f32_e32 v6, 1.0, v37
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v6
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v7
; SI-NEXT:    v_alignbit_b32 v6, v6, v7, 16
; SI-NEXT:    v_mul_f32_e32 v7, 1.0, v38
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v8
; SI-NEXT:    v_alignbit_b32 v7, v7, v8, 16
; SI-NEXT:    v_mul_f32_e32 v8, 1.0, v39
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v8
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v9
; SI-NEXT:    v_alignbit_b32 v8, v8, v9, 16
; SI-NEXT:    v_mul_f32_e32 v9, 1.0, v48
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v10
; SI-NEXT:    v_alignbit_b32 v9, v9, v10, 16
; SI-NEXT:    v_mul_f32_e32 v10, 1.0, v49
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v10
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v11
; SI-NEXT:    v_alignbit_b32 v10, v10, v11, 16
; SI-NEXT:    v_mul_f32_e32 v11, 1.0, v50
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v12
; SI-NEXT:    v_alignbit_b32 v11, v11, v12, 16
; SI-NEXT:    v_mul_f32_e32 v12, 1.0, v51
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v12
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v13
; SI-NEXT:    v_alignbit_b32 v12, v12, v13, 16
; SI-NEXT:    v_mul_f32_e32 v13, 1.0, v52
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v14
; SI-NEXT:    v_alignbit_b32 v13, v13, v14, 16
; SI-NEXT:    v_mul_f32_e32 v14, 1.0, v53
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v15
; SI-NEXT:    v_alignbit_b32 v14, v14, v15, 16
; SI-NEXT:    v_mul_f32_e32 v15, 1.0, v54
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v15
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v16
; SI-NEXT:    v_alignbit_b32 v15, v15, v16, 16
; SI-NEXT:    v_mul_f32_e32 v16, 1.0, v55
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v16
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v17
; SI-NEXT:    v_alignbit_b32 v16, v16, v17, 16
; SI-NEXT:    v_mul_f32_e32 v17, 1.0, v40
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v18
; SI-NEXT:    v_alignbit_b32 v17, v17, v18, 16
; SI-NEXT:    v_mul_f32_e32 v18, 1.0, v41
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v18
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v19
; SI-NEXT:    v_alignbit_b32 v18, v18, v19, 16
; SI-NEXT:    v_mul_f32_e32 v19, 1.0, v42
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v19
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v20
; SI-NEXT:    v_alignbit_b32 v19, v19, v20, 16
; SI-NEXT:    v_mul_f32_e32 v20, 1.0, v43
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v21
; SI-NEXT:    v_alignbit_b32 v20, v20, v21, 16
; SI-NEXT:    v_mul_f32_e32 v21, 1.0, v44
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v21
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v22
; SI-NEXT:    v_alignbit_b32 v21, v21, v22, 16
; SI-NEXT:    v_mul_f32_e32 v22, 1.0, v45
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v23
; SI-NEXT:    v_alignbit_b32 v22, v22, v23, 16
; SI-NEXT:    v_mul_f32_e32 v23, 1.0, v46
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v24
; SI-NEXT:    v_alignbit_b32 v23, v23, v24, 16
; SI-NEXT:    v_mul_f32_e32 v24, 1.0, v47
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v24
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v25
; SI-NEXT:    v_alignbit_b32 v24, v24, v25, 16
; SI-NEXT:    v_mul_f32_e32 v25, 1.0, v56
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v26
; SI-NEXT:    v_alignbit_b32 v25, v25, v26, 16
; SI-NEXT:    v_mul_f32_e32 v26, 1.0, v57
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v27
; SI-NEXT:    v_alignbit_b32 v26, v26, v27, 16
; SI-NEXT:    v_mul_f32_e32 v27, 1.0, v58
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v28
; SI-NEXT:    v_alignbit_b32 v27, v27, v28, 16
; SI-NEXT:    v_mul_f32_e32 v28, 1.0, v59
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v29
; SI-NEXT:    v_alignbit_b32 v28, v28, v29, 16
; SI-NEXT:    v_mul_f32_e32 v29, 1.0, v60
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v29
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v30
; SI-NEXT:    v_alignbit_b32 v29, v29, v30, 16
; SI-NEXT:    v_mul_f32_e32 v30, 1.0, v61
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v31
; SI-NEXT:    v_alignbit_b32 v30, v30, v31, 16
; SI-NEXT:    v_mul_f32_e32 v31, 1.0, v62
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_mul_f32_e32 v32, 1.0, v32
; SI-NEXT:    v_alignbit_b32 v31, v31, v32, 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v64bf16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB106_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 3
; VI-NEXT:    v_add_u16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v15, 3, v15
; VI-NEXT:    v_or_b32_e32 v15, v15, v33
; VI-NEXT:    v_add_u16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v14, 3, v14
; VI-NEXT:    v_or_b32_e32 v14, v14, v33
; VI-NEXT:    v_add_u16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v13, 3, v13
; VI-NEXT:    v_or_b32_e32 v13, v13, v33
; VI-NEXT:    v_add_u16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v12, 3, v12
; VI-NEXT:    v_or_b32_e32 v12, v12, v33
; VI-NEXT:    v_add_u16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v11, 3, v11
; VI-NEXT:    v_or_b32_e32 v11, v11, v33
; VI-NEXT:    v_add_u16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v10, 3, v10
; VI-NEXT:    v_or_b32_e32 v10, v10, v33
; VI-NEXT:    v_add_u16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v9, 3, v9
; VI-NEXT:    v_or_b32_e32 v9, v9, v33
; VI-NEXT:    v_add_u16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_e32 v8, v8, v33
; VI-NEXT:    v_add_u16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_e32 v7, v7, v33
; VI-NEXT:    v_add_u16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_e32 v6, v6, v33
; VI-NEXT:    v_add_u16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_e32 v5, v5, v33
; VI-NEXT:    v_add_u16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_e32 v4, v4, v33
; VI-NEXT:    v_add_u16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_e32 v3, v3, v33
; VI-NEXT:    v_add_u16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_e32 v2, v2, v33
; VI-NEXT:    v_add_u16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_e32 v1, v1, v33
; VI-NEXT:    v_add_u16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    v_or_b32_e32 v0, v0, v33
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v31, 3, v31
; VI-NEXT:    v_or_b32_e32 v31, v31, v33
; VI-NEXT:    v_add_u16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    v_or_b32_e32 v30, v30, v33
; VI-NEXT:    v_add_u16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    v_add_u16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v33
; VI-NEXT:    v_add_u16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v33
; VI-NEXT:    v_add_u16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v33
; VI-NEXT:    v_add_u16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v33
; VI-NEXT:    v_add_u16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v33
; VI-NEXT:    v_add_u16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v33
; VI-NEXT:    v_add_u16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v33
; VI-NEXT:    v_add_u16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v33
; VI-NEXT:    v_add_u16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_e32 v20, v20, v33
; VI-NEXT:    v_add_u16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_e32 v19, v19, v33
; VI-NEXT:    v_add_u16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    v_or_b32_e32 v18, v18, v33
; VI-NEXT:    v_add_u16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_add_u16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_e32 v17, v17, v33
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB106_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64i16_to_v64bf16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB106_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB106_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v64bf16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB106_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:  .LBB106_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define inreg <64 x bfloat> @bitcast_v64i16_to_v64bf16_scalar(<64 x i16> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v33, s30, 0
; SI-NEXT:    v_writelane_b32 v33, s31, 1
; SI-NEXT:    v_writelane_b32 v33, s34, 2
; SI-NEXT:    v_writelane_b32 v33, s35, 3
; SI-NEXT:    v_writelane_b32 v33, s36, 4
; SI-NEXT:    v_writelane_b32 v33, s37, 5
; SI-NEXT:    v_writelane_b32 v33, s38, 6
; SI-NEXT:    v_writelane_b32 v33, s39, 7
; SI-NEXT:    v_writelane_b32 v33, s48, 8
; SI-NEXT:    v_writelane_b32 v33, s49, 9
; SI-NEXT:    v_writelane_b32 v33, s50, 10
; SI-NEXT:    v_writelane_b32 v33, s51, 11
; SI-NEXT:    v_writelane_b32 v33, s52, 12
; SI-NEXT:    v_writelane_b32 v33, s53, 13
; SI-NEXT:    v_writelane_b32 v33, s54, 14
; SI-NEXT:    v_writelane_b32 v33, s55, 15
; SI-NEXT:    v_writelane_b32 v33, s64, 16
; SI-NEXT:    v_writelane_b32 v33, s65, 17
; SI-NEXT:    v_writelane_b32 v33, s66, 18
; SI-NEXT:    v_writelane_b32 v33, s67, 19
; SI-NEXT:    s_lshr_b32 s5, s16, 16
; SI-NEXT:    ; implicit-def: $vgpr34 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v33, s68, 20
; SI-NEXT:    s_lshr_b32 s6, s17, 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v34, s5, 0
; SI-NEXT:    v_writelane_b32 v33, s69, 21
; SI-NEXT:    s_lshr_b32 s8, s18, 16
; SI-NEXT:    v_writelane_b32 v34, s6, 1
; SI-NEXT:    v_writelane_b32 v33, s70, 22
; SI-NEXT:    s_lshr_b32 s90, s19, 16
; SI-NEXT:    v_writelane_b32 v34, s8, 2
; SI-NEXT:    v_writelane_b32 v33, s71, 23
; SI-NEXT:    s_lshr_b32 s91, s20, 16
; SI-NEXT:    v_writelane_b32 v34, s90, 3
; SI-NEXT:    v_writelane_b32 v33, s80, 24
; SI-NEXT:    s_lshr_b32 s92, s21, 16
; SI-NEXT:    v_writelane_b32 v34, s91, 4
; SI-NEXT:    v_writelane_b32 v33, s81, 25
; SI-NEXT:    s_lshr_b32 s93, s22, 16
; SI-NEXT:    v_writelane_b32 v34, s92, 5
; SI-NEXT:    v_writelane_b32 v33, s82, 26
; SI-NEXT:    s_lshr_b32 s94, s23, 16
; SI-NEXT:    v_writelane_b32 v34, s93, 6
; SI-NEXT:    v_writelane_b32 v33, s83, 27
; SI-NEXT:    s_lshr_b32 s95, s24, 16
; SI-NEXT:    v_writelane_b32 v34, s94, 7
; SI-NEXT:    v_writelane_b32 v33, s84, 28
; SI-NEXT:    s_lshr_b32 vcc_lo, s25, 16
; SI-NEXT:    v_writelane_b32 v34, s95, 8
; SI-NEXT:    v_writelane_b32 v33, s85, 29
; SI-NEXT:    s_lshr_b32 vcc_hi, s26, 16
; SI-NEXT:    v_writelane_b32 v34, vcc_lo, 9
; SI-NEXT:    v_writelane_b32 v33, s86, 30
; SI-NEXT:    s_lshr_b32 s30, s27, 16
; SI-NEXT:    v_writelane_b32 v34, vcc_hi, 10
; SI-NEXT:    v_writelane_b32 v33, s87, 31
; SI-NEXT:    s_lshr_b32 s31, s28, 16
; SI-NEXT:    v_writelane_b32 v34, s30, 11
; SI-NEXT:    v_writelane_b32 v33, s96, 32
; SI-NEXT:    s_lshr_b32 s34, s29, 16
; SI-NEXT:    v_readfirstlane_b32 s89, v17
; SI-NEXT:    v_readfirstlane_b32 s79, v16
; SI-NEXT:    v_readfirstlane_b32 s77, v15
; SI-NEXT:    v_readfirstlane_b32 s75, v14
; SI-NEXT:    v_readfirstlane_b32 s73, v13
; SI-NEXT:    v_readfirstlane_b32 s63, v12
; SI-NEXT:    v_readfirstlane_b32 s61, v11
; SI-NEXT:    v_readfirstlane_b32 s59, v10
; SI-NEXT:    v_readfirstlane_b32 s57, v9
; SI-NEXT:    v_readfirstlane_b32 s47, v8
; SI-NEXT:    v_readfirstlane_b32 s45, v7
; SI-NEXT:    v_readfirstlane_b32 s43, v6
; SI-NEXT:    v_readfirstlane_b32 s41, v5
; SI-NEXT:    v_readfirstlane_b32 s15, v4
; SI-NEXT:    v_readfirstlane_b32 s13, v3
; SI-NEXT:    v_readfirstlane_b32 s11, v2
; SI-NEXT:    v_readfirstlane_b32 s9, v1
; SI-NEXT:    v_readfirstlane_b32 s7, v0
; SI-NEXT:    v_writelane_b32 v34, s31, 12
; SI-NEXT:    v_writelane_b32 v33, s97, 33
; SI-NEXT:    s_lshr_b32 s88, s89, 16
; SI-NEXT:    s_lshr_b32 s78, s79, 16
; SI-NEXT:    s_lshr_b32 s76, s77, 16
; SI-NEXT:    s_lshr_b32 s74, s75, 16
; SI-NEXT:    s_lshr_b32 s72, s73, 16
; SI-NEXT:    s_lshr_b32 s62, s63, 16
; SI-NEXT:    s_lshr_b32 s60, s61, 16
; SI-NEXT:    s_lshr_b32 s58, s59, 16
; SI-NEXT:    s_lshr_b32 s56, s57, 16
; SI-NEXT:    s_lshr_b32 s46, s47, 16
; SI-NEXT:    s_lshr_b32 s44, s45, 16
; SI-NEXT:    s_lshr_b32 s42, s43, 16
; SI-NEXT:    s_lshr_b32 s40, s41, 16
; SI-NEXT:    s_lshr_b32 s14, s15, 16
; SI-NEXT:    s_lshr_b32 s12, s13, 16
; SI-NEXT:    s_lshr_b32 s10, s11, 16
; SI-NEXT:    s_lshr_b32 s36, s9, 16
; SI-NEXT:    s_lshr_b32 s35, s7, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v34, s34, 13
; SI-NEXT:    v_writelane_b32 v33, s98, 34
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v34, s35, 14
; SI-NEXT:    v_writelane_b32 v33, s99, 35
; SI-NEXT:    v_writelane_b32 v34, s36, 15
; SI-NEXT:    s_cbranch_scc0 .LBB107_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_lshl_b32 s4, s16, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 17
; SI-NEXT:    s_lshl_b32 s4, s5, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 16
; SI-NEXT:    s_lshl_b32 s4, s17, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 19
; SI-NEXT:    s_lshl_b32 s4, s6, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 18
; SI-NEXT:    s_lshl_b32 s4, s18, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 21
; SI-NEXT:    s_lshl_b32 s4, s8, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 20
; SI-NEXT:    s_lshl_b32 s4, s19, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 23
; SI-NEXT:    s_lshl_b32 s4, s90, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 22
; SI-NEXT:    s_lshl_b32 s4, s20, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 25
; SI-NEXT:    s_lshl_b32 s4, s91, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 24
; SI-NEXT:    s_lshl_b32 s4, s21, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 27
; SI-NEXT:    s_lshl_b32 s4, s92, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 26
; SI-NEXT:    s_lshl_b32 s4, s22, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 29
; SI-NEXT:    s_lshl_b32 s4, s93, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 28
; SI-NEXT:    s_lshl_b32 s4, s23, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 31
; SI-NEXT:    s_lshl_b32 s4, s94, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 30
; SI-NEXT:    s_lshl_b32 s4, s24, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 33
; SI-NEXT:    s_lshl_b32 s4, s95, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 32
; SI-NEXT:    s_lshl_b32 s4, s25, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 35
; SI-NEXT:    s_lshl_b32 s4, vcc_lo, 16
; SI-NEXT:    v_writelane_b32 v34, s4, 34
; SI-NEXT:    s_lshl_b32 s99, s26, 16
; SI-NEXT:    s_lshl_b32 s8, vcc_hi, 16
; SI-NEXT:    s_lshl_b32 s87, s27, 16
; SI-NEXT:    s_lshl_b32 s6, s30, 16
; SI-NEXT:    s_lshl_b32 s85, s28, 16
; SI-NEXT:    s_lshl_b32 s98, s31, 16
; SI-NEXT:    s_lshl_b32 s83, s29, 16
; SI-NEXT:    s_lshl_b32 s97, s34, 16
; SI-NEXT:    s_lshl_b32 s81, s7, 16
; SI-NEXT:    s_lshl_b32 s96, s35, 16
; SI-NEXT:    s_lshl_b32 s71, s9, 16
; SI-NEXT:    s_lshl_b32 s86, s36, 16
; SI-NEXT:    s_lshl_b32 s80, s11, 16
; SI-NEXT:    s_lshl_b32 s84, s10, 16
; SI-NEXT:    s_lshl_b32 s67, s13, 16
; SI-NEXT:    s_lshl_b32 s82, s12, 16
; SI-NEXT:    s_lshl_b32 s65, s15, 16
; SI-NEXT:    s_lshl_b32 s69, s14, 16
; SI-NEXT:    s_lshl_b32 s55, s41, 16
; SI-NEXT:    s_lshl_b32 s70, s40, 16
; SI-NEXT:    s_lshl_b32 s53, s43, 16
; SI-NEXT:    s_lshl_b32 s68, s42, 16
; SI-NEXT:    s_lshl_b32 s51, s45, 16
; SI-NEXT:    s_lshl_b32 s66, s44, 16
; SI-NEXT:    s_lshl_b32 s49, s47, 16
; SI-NEXT:    s_lshl_b32 s64, s46, 16
; SI-NEXT:    s_lshl_b32 s39, s57, 16
; SI-NEXT:    s_lshl_b32 s54, s56, 16
; SI-NEXT:    s_lshl_b32 s37, s59, 16
; SI-NEXT:    s_lshl_b32 s52, s58, 16
; SI-NEXT:    s_lshl_b32 s35, s61, 16
; SI-NEXT:    s_lshl_b32 s50, s60, 16
; SI-NEXT:    s_lshl_b32 s31, s63, 16
; SI-NEXT:    s_lshl_b32 s48, s62, 16
; SI-NEXT:    s_lshl_b32 s95, s73, 16
; SI-NEXT:    s_lshl_b32 s38, s72, 16
; SI-NEXT:    s_lshl_b32 s34, s75, 16
; SI-NEXT:    s_lshl_b32 s36, s74, 16
; SI-NEXT:    s_lshl_b32 s92, s77, 16
; SI-NEXT:    s_lshl_b32 s30, s76, 16
; SI-NEXT:    s_lshl_b32 s93, s79, 16
; SI-NEXT:    s_lshl_b32 s94, s78, 16
; SI-NEXT:    s_lshl_b32 s90, s89, 16
; SI-NEXT:    s_lshl_b32 s91, s88, 16
; SI-NEXT:    s_mov_b64 s[4:5], 0
; SI-NEXT:    s_branch .LBB107_3
; SI-NEXT:  .LBB107_2:
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    s_mov_b64 s[4:5], -1
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr99
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr87
; SI-NEXT:    ; implicit-def: $sgpr85
; SI-NEXT:    ; implicit-def: $sgpr98
; SI-NEXT:    ; implicit-def: $sgpr83
; SI-NEXT:    ; implicit-def: $sgpr97
; SI-NEXT:    ; implicit-def: $sgpr81
; SI-NEXT:    ; implicit-def: $sgpr96
; SI-NEXT:    ; implicit-def: $sgpr71
; SI-NEXT:    ; implicit-def: $sgpr86
; SI-NEXT:    ; implicit-def: $sgpr80
; SI-NEXT:    ; implicit-def: $sgpr84
; SI-NEXT:    ; implicit-def: $sgpr67
; SI-NEXT:    ; implicit-def: $sgpr82
; SI-NEXT:    ; implicit-def: $sgpr65
; SI-NEXT:    ; implicit-def: $sgpr69
; SI-NEXT:    ; implicit-def: $sgpr55
; SI-NEXT:    ; implicit-def: $sgpr70
; SI-NEXT:    ; implicit-def: $sgpr53
; SI-NEXT:    ; implicit-def: $sgpr68
; SI-NEXT:    ; implicit-def: $sgpr51
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    ; implicit-def: $sgpr49
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr39
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr37
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr35
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr31
; SI-NEXT:    ; implicit-def: $sgpr48
; SI-NEXT:    ; implicit-def: $sgpr95
; SI-NEXT:    ; implicit-def: $sgpr38
; SI-NEXT:    ; implicit-def: $sgpr34
; SI-NEXT:    ; implicit-def: $sgpr36
; SI-NEXT:    ; implicit-def: $sgpr92
; SI-NEXT:    ; implicit-def: $sgpr30
; SI-NEXT:    ; implicit-def: $sgpr93
; SI-NEXT:    ; implicit-def: $sgpr94
; SI-NEXT:    ; implicit-def: $sgpr90
; SI-NEXT:    ; implicit-def: $sgpr91
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; kill: killed $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:  .LBB107_3: ; %Flow
; SI-NEXT:    s_andn2_b64 vcc, exec, s[4:5]
; SI-NEXT:    s_mov_b32 s4, s90
; SI-NEXT:    s_mov_b32 s5, s93
; SI-NEXT:    s_mov_b32 s90, s92
; SI-NEXT:    s_mov_b32 s92, s34
; SI-NEXT:    s_mov_b32 s93, s95
; SI-NEXT:    s_mov_b32 s95, s31
; SI-NEXT:    s_mov_b32 s31, s35
; SI-NEXT:    s_mov_b32 s34, s36
; SI-NEXT:    s_mov_b32 s35, s37
; SI-NEXT:    s_mov_b32 s36, s38
; SI-NEXT:    s_mov_b32 s37, s39
; SI-NEXT:    s_mov_b32 s38, s48
; SI-NEXT:    s_mov_b32 s39, s49
; SI-NEXT:    s_mov_b32 s48, s50
; SI-NEXT:    s_mov_b32 s49, s51
; SI-NEXT:    s_mov_b32 s50, s52
; SI-NEXT:    s_mov_b32 s51, s53
; SI-NEXT:    s_mov_b32 s52, s54
; SI-NEXT:    s_mov_b32 s53, s55
; SI-NEXT:    s_mov_b32 s54, s64
; SI-NEXT:    s_mov_b32 s55, s65
; SI-NEXT:    s_mov_b32 s64, s66
; SI-NEXT:    s_mov_b32 s65, s67
; SI-NEXT:    s_mov_b32 s66, s68
; SI-NEXT:    s_mov_b32 s67, s80
; SI-NEXT:    s_mov_b32 s68, s70
; SI-NEXT:    s_mov_b32 s70, s82
; SI-NEXT:    s_mov_b32 s80, s6
; SI-NEXT:    s_mov_b32 s82, s8
; SI-NEXT:    s_cbranch_vccnz .LBB107_5
; SI-NEXT:  ; %bb.4: ; %cmp.true
; SI-NEXT:    s_add_i32 s9, s9, 3
; SI-NEXT:    v_readlane_b32 s6, v34, 15
; SI-NEXT:    s_add_i32 s11, s11, 3
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s8, s6, 16
; SI-NEXT:    s_add_i32 s7, s7, 3
; SI-NEXT:    v_readlane_b32 s6, v34, 14
; SI-NEXT:    s_add_i32 s13, s13, 3
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s10, s10, 16
; SI-NEXT:    s_or_b32 s8, s8, s9
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s6, s6, 16
; SI-NEXT:    s_add_i32 s29, s29, 3
; SI-NEXT:    v_readlane_b32 s9, v34, 13
; SI-NEXT:    s_add_i32 s15, s15, 3
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s12, s12, 16
; SI-NEXT:    s_or_b32 s10, s10, s11
; SI-NEXT:    s_or_b32 s6, s6, s7
; SI-NEXT:    s_and_b32 s7, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_add_i32 s28, s28, 3
; SI-NEXT:    v_readlane_b32 s11, v34, 12
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s14, s14, 16
; SI-NEXT:    s_or_b32 s12, s12, s13
; SI-NEXT:    s_or_b32 s7, s9, s7
; SI-NEXT:    s_and_b32 s9, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_add_i32 s27, s27, 3
; SI-NEXT:    v_readlane_b32 s13, v34, 11
; SI-NEXT:    s_or_b32 s14, s14, s15
; SI-NEXT:    s_or_b32 s9, s11, s9
; SI-NEXT:    s_and_b32 s11, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s13, s13, 16
; SI-NEXT:    s_add_i32 s26, s26, 3
; SI-NEXT:    v_readlane_b32 s15, v34, 10
; SI-NEXT:    s_or_b32 s11, s13, s11
; SI-NEXT:    s_and_b32 s13, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    s_or_b32 s13, s15, s13
; SI-NEXT:    s_and_b32 s15, s25, 0xffff
; SI-NEXT:    v_readlane_b32 s25, v34, 9
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s15, s25, s15
; SI-NEXT:    s_add_i32 s24, s24, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 8
; SI-NEXT:    s_and_b32 s24, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s24, s25, s24
; SI-NEXT:    s_add_i32 s23, s23, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 7
; SI-NEXT:    s_and_b32 s23, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s23, s25, s23
; SI-NEXT:    s_add_i32 s22, s22, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 6
; SI-NEXT:    s_and_b32 s22, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s22, s25, s22
; SI-NEXT:    s_add_i32 s21, s21, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 5
; SI-NEXT:    s_and_b32 s21, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s21, s25, s21
; SI-NEXT:    s_add_i32 s20, s20, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 4
; SI-NEXT:    s_and_b32 s20, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s20, s25, s20
; SI-NEXT:    s_add_i32 s19, s19, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 3
; SI-NEXT:    s_and_b32 s19, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s19, s25, s19
; SI-NEXT:    s_add_i32 s18, s18, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 2
; SI-NEXT:    s_and_b32 s18, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s18, s25, s18
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 1
; SI-NEXT:    s_and_b32 s17, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s17, s25, s17
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s25, v34, 0
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s25, 16
; SI-NEXT:    s_or_b32 s16, s25, s16
; SI-NEXT:    s_add_i32 s16, s16, 0x30000
; SI-NEXT:    s_and_b32 s25, s16, 0xffff0000
; SI-NEXT:    s_add_i32 s17, s17, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s25, 16
; SI-NEXT:    s_lshl_b32 s16, s16, 16
; SI-NEXT:    v_writelane_b32 v34, s16, 17
; SI-NEXT:    s_and_b32 s16, s17, 0xffff0000
; SI-NEXT:    s_add_i32 s18, s18, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 18
; SI-NEXT:    s_lshl_b32 s16, s17, 16
; SI-NEXT:    v_writelane_b32 v34, s16, 19
; SI-NEXT:    s_and_b32 s16, s18, 0xffff0000
; SI-NEXT:    s_add_i32 s19, s19, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 20
; SI-NEXT:    s_lshl_b32 s16, s18, 16
; SI-NEXT:    v_writelane_b32 v34, s16, 21
; SI-NEXT:    s_and_b32 s16, s19, 0xffff0000
; SI-NEXT:    s_add_i32 s20, s20, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 22
; SI-NEXT:    s_lshl_b32 s16, s19, 16
; SI-NEXT:    v_writelane_b32 v34, s16, 23
; SI-NEXT:    s_and_b32 s16, s20, 0xffff0000
; SI-NEXT:    s_add_i32 s21, s21, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 24
; SI-NEXT:    s_lshl_b32 s16, s20, 16
; SI-NEXT:    v_writelane_b32 v34, s16, 25
; SI-NEXT:    s_and_b32 s16, s21, 0xffff0000
; SI-NEXT:    s_add_i32 s22, s22, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 26
; SI-NEXT:    s_lshl_b32 s16, s21, 16
; SI-NEXT:    v_writelane_b32 v34, s16, 27
; SI-NEXT:    s_and_b32 s16, s22, 0xffff0000
; SI-NEXT:    s_add_i32 s23, s23, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 28
; SI-NEXT:    s_lshl_b32 s16, s22, 16
; SI-NEXT:    s_add_i32 s89, s89, 3
; SI-NEXT:    v_writelane_b32 v34, s16, 29
; SI-NEXT:    s_and_b32 s16, s23, 0xffff0000
; SI-NEXT:    s_and_b32 s4, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s88, 16
; SI-NEXT:    s_add_i32 s79, s79, 3
; SI-NEXT:    s_add_i32 s77, s77, 3
; SI-NEXT:    s_add_i32 s75, s75, 3
; SI-NEXT:    s_add_i32 s73, s73, 3
; SI-NEXT:    s_add_i32 s63, s63, 3
; SI-NEXT:    s_add_i32 s61, s61, 3
; SI-NEXT:    s_add_i32 s59, s59, 3
; SI-NEXT:    s_add_i32 s57, s57, 3
; SI-NEXT:    s_add_i32 s47, s47, 3
; SI-NEXT:    s_add_i32 s45, s45, 3
; SI-NEXT:    s_add_i32 s43, s43, 3
; SI-NEXT:    s_add_i32 s41, s41, 3
; SI-NEXT:    s_add_i32 s24, s24, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 30
; SI-NEXT:    s_lshl_b32 s16, s23, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    s_and_b32 s5, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s78, s78, 16
; SI-NEXT:    s_and_b32 s77, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s76, s76, 16
; SI-NEXT:    s_and_b32 s75, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s74, s74, 16
; SI-NEXT:    s_and_b32 s73, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s72, s72, 16
; SI-NEXT:    s_and_b32 s63, s63, 0xffff
; SI-NEXT:    s_lshl_b32 s62, s62, 16
; SI-NEXT:    s_and_b32 s61, s61, 0xffff
; SI-NEXT:    s_lshl_b32 s60, s60, 16
; SI-NEXT:    s_and_b32 s59, s59, 0xffff
; SI-NEXT:    s_lshl_b32 s58, s58, 16
; SI-NEXT:    s_and_b32 s57, s57, 0xffff
; SI-NEXT:    s_lshl_b32 s56, s56, 16
; SI-NEXT:    s_and_b32 s47, s47, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s44, 16
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s42, s42, 16
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s40, s40, 16
; SI-NEXT:    v_writelane_b32 v34, s16, 31
; SI-NEXT:    s_and_b32 s16, s24, 0xffff0000
; SI-NEXT:    s_or_b32 s5, s78, s5
; SI-NEXT:    s_or_b32 s76, s76, s77
; SI-NEXT:    s_or_b32 s74, s74, s75
; SI-NEXT:    s_or_b32 s72, s72, s73
; SI-NEXT:    s_or_b32 s62, s62, s63
; SI-NEXT:    s_or_b32 s60, s60, s61
; SI-NEXT:    s_or_b32 s58, s58, s59
; SI-NEXT:    s_or_b32 s56, s56, s57
; SI-NEXT:    s_or_b32 s46, s46, s47
; SI-NEXT:    s_or_b32 s44, s44, s45
; SI-NEXT:    s_or_b32 s42, s42, s43
; SI-NEXT:    s_or_b32 s40, s40, s41
; SI-NEXT:    s_add_i32 s15, s15, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 32
; SI-NEXT:    s_lshl_b32 s16, s24, 16
; SI-NEXT:    s_add_i32 s4, s4, 0x30000
; SI-NEXT:    s_add_i32 s5, s5, 0x30000
; SI-NEXT:    s_add_i32 s76, s76, 0x30000
; SI-NEXT:    s_add_i32 s74, s74, 0x30000
; SI-NEXT:    s_add_i32 s72, s72, 0x30000
; SI-NEXT:    s_add_i32 s62, s62, 0x30000
; SI-NEXT:    s_add_i32 s60, s60, 0x30000
; SI-NEXT:    s_add_i32 s58, s58, 0x30000
; SI-NEXT:    s_add_i32 s56, s56, 0x30000
; SI-NEXT:    s_add_i32 s46, s46, 0x30000
; SI-NEXT:    s_add_i32 s44, s44, 0x30000
; SI-NEXT:    s_add_i32 s42, s42, 0x30000
; SI-NEXT:    s_add_i32 s40, s40, 0x30000
; SI-NEXT:    s_add_i32 s14, s14, 0x30000
; SI-NEXT:    s_add_i32 s12, s12, 0x30000
; SI-NEXT:    s_add_i32 s10, s10, 0x30000
; SI-NEXT:    s_add_i32 s8, s8, 0x30000
; SI-NEXT:    s_add_i32 s6, s6, 0x30000
; SI-NEXT:    s_add_i32 s7, s7, 0x30000
; SI-NEXT:    s_add_i32 s9, s9, 0x30000
; SI-NEXT:    s_add_i32 s11, s11, 0x30000
; SI-NEXT:    s_add_i32 s13, s13, 0x30000
; SI-NEXT:    v_writelane_b32 v34, s16, 33
; SI-NEXT:    s_and_b32 s16, s15, 0xffff0000
; SI-NEXT:    v_writelane_b32 v34, s16, 34
; SI-NEXT:    s_lshl_b32 s15, s15, 16
; SI-NEXT:    s_and_b32 s82, s13, 0xffff0000
; SI-NEXT:    s_lshl_b32 s99, s13, 16
; SI-NEXT:    s_and_b32 s80, s11, 0xffff0000
; SI-NEXT:    s_lshl_b32 s87, s11, 16
; SI-NEXT:    s_and_b32 s98, s9, 0xffff0000
; SI-NEXT:    s_lshl_b32 s85, s9, 16
; SI-NEXT:    s_and_b32 s97, s7, 0xffff0000
; SI-NEXT:    s_lshl_b32 s83, s7, 16
; SI-NEXT:    s_and_b32 s96, s6, 0xffff0000
; SI-NEXT:    s_lshl_b32 s81, s6, 16
; SI-NEXT:    s_and_b32 s86, s8, 0xffff0000
; SI-NEXT:    s_lshl_b32 s71, s8, 16
; SI-NEXT:    s_and_b32 s84, s10, 0xffff0000
; SI-NEXT:    s_lshl_b32 s67, s10, 16
; SI-NEXT:    s_and_b32 s70, s12, 0xffff0000
; SI-NEXT:    s_lshl_b32 s65, s12, 16
; SI-NEXT:    s_and_b32 s69, s14, 0xffff0000
; SI-NEXT:    s_lshl_b32 s55, s14, 16
; SI-NEXT:    s_and_b32 s68, s40, 0xffff0000
; SI-NEXT:    s_lshl_b32 s53, s40, 16
; SI-NEXT:    s_and_b32 s66, s42, 0xffff0000
; SI-NEXT:    s_lshl_b32 s51, s42, 16
; SI-NEXT:    s_and_b32 s64, s44, 0xffff0000
; SI-NEXT:    s_lshl_b32 s49, s44, 16
; SI-NEXT:    s_and_b32 s54, s46, 0xffff0000
; SI-NEXT:    s_lshl_b32 s39, s46, 16
; SI-NEXT:    s_and_b32 s52, s56, 0xffff0000
; SI-NEXT:    s_lshl_b32 s37, s56, 16
; SI-NEXT:    s_and_b32 s50, s58, 0xffff0000
; SI-NEXT:    s_lshl_b32 s35, s58, 16
; SI-NEXT:    s_and_b32 s48, s60, 0xffff0000
; SI-NEXT:    s_lshl_b32 s31, s60, 16
; SI-NEXT:    s_and_b32 s38, s62, 0xffff0000
; SI-NEXT:    s_lshl_b32 s95, s62, 16
; SI-NEXT:    s_and_b32 s36, s72, 0xffff0000
; SI-NEXT:    s_lshl_b32 s93, s72, 16
; SI-NEXT:    s_and_b32 s34, s74, 0xffff0000
; SI-NEXT:    s_lshl_b32 s92, s74, 16
; SI-NEXT:    s_and_b32 s30, s76, 0xffff0000
; SI-NEXT:    s_lshl_b32 s90, s76, 16
; SI-NEXT:    s_and_b32 s94, s5, 0xffff0000
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_and_b32 s91, s4, 0xffff0000
; SI-NEXT:    s_lshl_b32 s4, s4, 16
; SI-NEXT:    v_writelane_b32 v34, s15, 35
; SI-NEXT:  .LBB107_5: ; %end
; SI-NEXT:    v_readlane_b32 s6, v34, 16
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 17
; SI-NEXT:    v_lshrrev_b32_e32 v1, 16, v0
; SI-NEXT:    v_mul_f32_e64 v0, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[0:1], v[0:1], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 18
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 19
; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
; SI-NEXT:    v_mul_f32_e64 v1, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[1:2], v[1:2], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 20
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 21
; SI-NEXT:    v_lshrrev_b32_e32 v3, 16, v2
; SI-NEXT:    v_mul_f32_e64 v2, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[2:3], v[2:3], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 22
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 23
; SI-NEXT:    v_lshrrev_b32_e32 v4, 16, v3
; SI-NEXT:    v_mul_f32_e64 v3, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[3:4], v[3:4], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 24
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 25
; SI-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
; SI-NEXT:    v_mul_f32_e64 v4, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[4:5], v[4:5], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 26
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 27
; SI-NEXT:    v_lshrrev_b32_e32 v6, 16, v5
; SI-NEXT:    v_mul_f32_e64 v5, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[5:6], v[5:6], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 28
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 29
; SI-NEXT:    v_lshrrev_b32_e32 v7, 16, v6
; SI-NEXT:    v_mul_f32_e64 v6, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[6:7], v[6:7], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 30
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 31
; SI-NEXT:    v_lshrrev_b32_e32 v8, 16, v7
; SI-NEXT:    v_mul_f32_e64 v7, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[7:8], v[7:8], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 32
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 33
; SI-NEXT:    v_lshrrev_b32_e32 v9, 16, v8
; SI-NEXT:    v_mul_f32_e64 v8, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[8:9], v[8:9], 16
; SI-NEXT:    v_readlane_b32 s6, v34, 34
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s6
; SI-NEXT:    v_readlane_b32 s6, v34, 35
; SI-NEXT:    v_lshrrev_b32_e32 v10, 16, v9
; SI-NEXT:    v_mul_f32_e64 v9, 1.0, s6
; SI-NEXT:    v_lshr_b64 v[9:10], v[9:10], 16
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s82
; SI-NEXT:    v_lshrrev_b32_e32 v11, 16, v10
; SI-NEXT:    v_mul_f32_e64 v10, 1.0, s99
; SI-NEXT:    v_lshr_b64 v[10:11], v[10:11], 16
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s80
; SI-NEXT:    v_lshrrev_b32_e32 v12, 16, v11
; SI-NEXT:    v_mul_f32_e64 v11, 1.0, s87
; SI-NEXT:    v_lshr_b64 v[11:12], v[11:12], 16
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s98
; SI-NEXT:    v_lshrrev_b32_e32 v13, 16, v12
; SI-NEXT:    v_mul_f32_e64 v12, 1.0, s85
; SI-NEXT:    v_lshr_b64 v[12:13], v[12:13], 16
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s97
; SI-NEXT:    v_lshrrev_b32_e32 v14, 16, v13
; SI-NEXT:    v_mul_f32_e64 v13, 1.0, s83
; SI-NEXT:    v_lshr_b64 v[13:14], v[13:14], 16
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s96
; SI-NEXT:    v_lshrrev_b32_e32 v15, 16, v14
; SI-NEXT:    v_mul_f32_e64 v14, 1.0, s81
; SI-NEXT:    v_lshr_b64 v[14:15], v[14:15], 16
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s86
; SI-NEXT:    v_lshrrev_b32_e32 v16, 16, v15
; SI-NEXT:    v_mul_f32_e64 v15, 1.0, s71
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s84
; SI-NEXT:    v_lshrrev_b32_e32 v17, 16, v16
; SI-NEXT:    v_mul_f32_e64 v16, 1.0, s67
; SI-NEXT:    v_lshr_b64 v[16:17], v[16:17], 16
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s70
; SI-NEXT:    v_lshrrev_b32_e32 v18, 16, v17
; SI-NEXT:    v_mul_f32_e64 v17, 1.0, s65
; SI-NEXT:    v_lshr_b64 v[17:18], v[17:18], 16
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s69
; SI-NEXT:    v_lshrrev_b32_e32 v19, 16, v18
; SI-NEXT:    v_mul_f32_e64 v18, 1.0, s55
; SI-NEXT:    v_lshr_b64 v[18:19], v[18:19], 16
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s68
; SI-NEXT:    v_lshrrev_b32_e32 v20, 16, v19
; SI-NEXT:    v_mul_f32_e64 v19, 1.0, s53
; SI-NEXT:    v_lshr_b64 v[19:20], v[19:20], 16
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s66
; SI-NEXT:    v_lshrrev_b32_e32 v21, 16, v20
; SI-NEXT:    v_mul_f32_e64 v20, 1.0, s51
; SI-NEXT:    v_lshr_b64 v[20:21], v[20:21], 16
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s64
; SI-NEXT:    v_lshrrev_b32_e32 v22, 16, v21
; SI-NEXT:    v_mul_f32_e64 v21, 1.0, s49
; SI-NEXT:    v_lshr_b64 v[21:22], v[21:22], 16
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s54
; SI-NEXT:    v_lshrrev_b32_e32 v23, 16, v22
; SI-NEXT:    v_mul_f32_e64 v22, 1.0, s39
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s52
; SI-NEXT:    v_lshrrev_b32_e32 v24, 16, v23
; SI-NEXT:    v_mul_f32_e64 v23, 1.0, s37
; SI-NEXT:    v_lshr_b64 v[23:24], v[23:24], 16
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s50
; SI-NEXT:    v_lshrrev_b32_e32 v25, 16, v24
; SI-NEXT:    v_mul_f32_e64 v24, 1.0, s35
; SI-NEXT:    v_lshr_b64 v[24:25], v[24:25], 16
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s48
; SI-NEXT:    v_lshrrev_b32_e32 v26, 16, v25
; SI-NEXT:    v_mul_f32_e64 v25, 1.0, s31
; SI-NEXT:    v_lshr_b64 v[25:26], v[25:26], 16
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s38
; SI-NEXT:    v_lshrrev_b32_e32 v27, 16, v26
; SI-NEXT:    v_mul_f32_e64 v26, 1.0, s95
; SI-NEXT:    v_lshr_b64 v[26:27], v[26:27], 16
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s36
; SI-NEXT:    v_lshrrev_b32_e32 v28, 16, v27
; SI-NEXT:    v_mul_f32_e64 v27, 1.0, s93
; SI-NEXT:    v_lshr_b64 v[27:28], v[27:28], 16
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s34
; SI-NEXT:    v_lshrrev_b32_e32 v29, 16, v28
; SI-NEXT:    v_mul_f32_e64 v28, 1.0, s92
; SI-NEXT:    v_lshr_b64 v[28:29], v[28:29], 16
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s30
; SI-NEXT:    v_lshrrev_b32_e32 v30, 16, v29
; SI-NEXT:    v_mul_f32_e64 v29, 1.0, s90
; SI-NEXT:    v_lshr_b64 v[29:30], v[29:30], 16
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s94
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v30
; SI-NEXT:    v_mul_f32_e64 v30, 1.0, s5
; SI-NEXT:    v_lshr_b64 v[30:31], v[30:31], 16
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s91
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v31
; SI-NEXT:    v_mul_f32_e64 v31, 1.0, s4
; SI-NEXT:    v_lshr_b64 v[31:32], v[31:32], 16
; SI-NEXT:    v_readlane_b32 s99, v33, 35
; SI-NEXT:    v_readlane_b32 s98, v33, 34
; SI-NEXT:    v_readlane_b32 s97, v33, 33
; SI-NEXT:    v_readlane_b32 s96, v33, 32
; SI-NEXT:    v_readlane_b32 s87, v33, 31
; SI-NEXT:    v_readlane_b32 s86, v33, 30
; SI-NEXT:    v_readlane_b32 s85, v33, 29
; SI-NEXT:    v_readlane_b32 s84, v33, 28
; SI-NEXT:    v_readlane_b32 s83, v33, 27
; SI-NEXT:    v_readlane_b32 s82, v33, 26
; SI-NEXT:    v_readlane_b32 s81, v33, 25
; SI-NEXT:    v_readlane_b32 s80, v33, 24
; SI-NEXT:    v_readlane_b32 s71, v33, 23
; SI-NEXT:    v_readlane_b32 s70, v33, 22
; SI-NEXT:    v_readlane_b32 s69, v33, 21
; SI-NEXT:    v_readlane_b32 s68, v33, 20
; SI-NEXT:    v_readlane_b32 s67, v33, 19
; SI-NEXT:    v_readlane_b32 s66, v33, 18
; SI-NEXT:    v_readlane_b32 s65, v33, 17
; SI-NEXT:    v_readlane_b32 s64, v33, 16
; SI-NEXT:    v_readlane_b32 s55, v33, 15
; SI-NEXT:    v_readlane_b32 s54, v33, 14
; SI-NEXT:    v_readlane_b32 s53, v33, 13
; SI-NEXT:    v_readlane_b32 s52, v33, 12
; SI-NEXT:    v_readlane_b32 s51, v33, 11
; SI-NEXT:    v_readlane_b32 s50, v33, 10
; SI-NEXT:    v_readlane_b32 s49, v33, 9
; SI-NEXT:    v_readlane_b32 s48, v33, 8
; SI-NEXT:    v_readlane_b32 s39, v33, 7
; SI-NEXT:    v_readlane_b32 s38, v33, 6
; SI-NEXT:    v_readlane_b32 s37, v33, 5
; SI-NEXT:    v_readlane_b32 s36, v33, 4
; SI-NEXT:    v_readlane_b32 s35, v33, 3
; SI-NEXT:    v_readlane_b32 s34, v33, 2
; SI-NEXT:    v_readlane_b32 s31, v33, 1
; SI-NEXT:    v_readlane_b32 s30, v33, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s34, 2
; VI-NEXT:    v_writelane_b32 v32, s35, 3
; VI-NEXT:    v_writelane_b32 v32, s36, 4
; VI-NEXT:    v_writelane_b32 v32, s37, 5
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s38, 6
; VI-NEXT:    v_readfirstlane_b32 s8, v17
; VI-NEXT:    v_readfirstlane_b32 s9, v16
; VI-NEXT:    v_readfirstlane_b32 s10, v15
; VI-NEXT:    v_readfirstlane_b32 s11, v14
; VI-NEXT:    v_readfirstlane_b32 s12, v13
; VI-NEXT:    v_readfirstlane_b32 s13, v12
; VI-NEXT:    v_readfirstlane_b32 s14, v11
; VI-NEXT:    v_readfirstlane_b32 s15, v10
; VI-NEXT:    v_readfirstlane_b32 s40, v9
; VI-NEXT:    v_readfirstlane_b32 s41, v8
; VI-NEXT:    v_readfirstlane_b32 s42, v7
; VI-NEXT:    v_readfirstlane_b32 s43, v6
; VI-NEXT:    v_readfirstlane_b32 s44, v5
; VI-NEXT:    v_readfirstlane_b32 s45, v4
; VI-NEXT:    v_readfirstlane_b32 s46, v3
; VI-NEXT:    v_readfirstlane_b32 s47, v2
; VI-NEXT:    v_readfirstlane_b32 s6, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s7, v0
; VI-NEXT:    v_writelane_b32 v32, s39, 7
; VI-NEXT:    s_cbranch_scc0 .LBB107_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB107_3
; VI-NEXT:  .LBB107_2: ; %cmp.true
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    s_add_i32 s5, s47, 3
; VI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_and_b32 s56, s45, 0xffff0000
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:    s_and_b32 s57, s44, 0xffff0000
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_and_b32 s58, s43, 0xffff0000
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_and_b32 s59, s42, 0xffff0000
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_and_b32 s60, s41, 0xffff0000
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_and_b32 s61, s40, 0xffff0000
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_and_b32 s62, s15, 0xffff0000
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_and_b32 s63, s14, 0xffff0000
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_and_b32 s72, s13, 0xffff0000
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_and_b32 s73, s12, 0xffff0000
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_and_b32 s74, s11, 0xffff0000
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_and_b32 s75, s10, 0xffff0000
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_and_b32 s76, s9, 0xffff0000
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_and_b32 s77, s8, 0xffff0000
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_and_b32 s78, s16, 0xffff0000
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    s_and_b32 s79, s17, 0xffff0000
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_and_b32 s88, s18, 0xffff0000
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_and_b32 s89, s19, 0xffff0000
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_and_b32 s90, s20, 0xffff0000
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_and_b32 s91, s21, 0xffff0000
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_and_b32 vcc_lo, s22, 0xffff0000
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_and_b32 vcc_hi, s23, 0xffff0000
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_and_b32 s30, s24, 0xffff0000
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_and_b32 s31, s25, 0xffff0000
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_and_b32 s34, s26, 0xffff0000
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_and_b32 s35, s27, 0xffff0000
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_and_b32 s36, s28, 0xffff0000
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_and_b32 s37, s29, 0xffff0000
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_and_b32 s38, s7, 0xffff0000
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_and_b32 s39, s6, 0xffff0000
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_and_b32 s6, s6, 0xffff
; VI-NEXT:    s_and_b32 s7, s7, 0xffff
; VI-NEXT:    s_and_b32 s29, s29, 0xffff
; VI-NEXT:    s_and_b32 s28, s28, 0xffff
; VI-NEXT:    s_and_b32 s27, s27, 0xffff
; VI-NEXT:    s_and_b32 s26, s26, 0xffff
; VI-NEXT:    s_and_b32 s25, s25, 0xffff
; VI-NEXT:    s_and_b32 s24, s24, 0xffff
; VI-NEXT:    s_and_b32 s23, s23, 0xffff
; VI-NEXT:    s_and_b32 s22, s22, 0xffff
; VI-NEXT:    s_and_b32 s21, s21, 0xffff
; VI-NEXT:    s_and_b32 s20, s20, 0xffff
; VI-NEXT:    s_and_b32 s19, s19, 0xffff
; VI-NEXT:    s_and_b32 s18, s18, 0xffff
; VI-NEXT:    s_and_b32 s17, s17, 0xffff
; VI-NEXT:    s_and_b32 s16, s16, 0xffff
; VI-NEXT:    s_and_b32 s8, s8, 0xffff
; VI-NEXT:    s_and_b32 s9, s9, 0xffff
; VI-NEXT:    s_and_b32 s10, s10, 0xffff
; VI-NEXT:    s_and_b32 s11, s11, 0xffff
; VI-NEXT:    s_and_b32 s12, s12, 0xffff
; VI-NEXT:    s_and_b32 s13, s13, 0xffff
; VI-NEXT:    s_and_b32 s14, s14, 0xffff
; VI-NEXT:    s_and_b32 s15, s15, 0xffff
; VI-NEXT:    s_and_b32 s40, s40, 0xffff
; VI-NEXT:    s_and_b32 s41, s41, 0xffff
; VI-NEXT:    s_and_b32 s42, s42, 0xffff
; VI-NEXT:    s_and_b32 s43, s43, 0xffff
; VI-NEXT:    s_and_b32 s44, s44, 0xffff
; VI-NEXT:    s_and_b32 s45, s45, 0xffff
; VI-NEXT:    s_and_b32 s46, s46, 0xffff
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s6, s39, s6
; VI-NEXT:    s_or_b32 s7, s38, s7
; VI-NEXT:    s_or_b32 s29, s37, s29
; VI-NEXT:    s_or_b32 s28, s36, s28
; VI-NEXT:    s_or_b32 s27, s35, s27
; VI-NEXT:    s_or_b32 s26, s34, s26
; VI-NEXT:    s_or_b32 s25, s31, s25
; VI-NEXT:    s_or_b32 s24, s30, s24
; VI-NEXT:    s_or_b32 s23, vcc_hi, s23
; VI-NEXT:    s_or_b32 s22, vcc_lo, s22
; VI-NEXT:    s_or_b32 s21, s91, s21
; VI-NEXT:    s_or_b32 s20, s90, s20
; VI-NEXT:    s_or_b32 s19, s89, s19
; VI-NEXT:    s_or_b32 s18, s88, s18
; VI-NEXT:    s_or_b32 s17, s79, s17
; VI-NEXT:    s_or_b32 s16, s78, s16
; VI-NEXT:    s_or_b32 s8, s77, s8
; VI-NEXT:    s_or_b32 s9, s76, s9
; VI-NEXT:    s_or_b32 s10, s75, s10
; VI-NEXT:    s_or_b32 s11, s74, s11
; VI-NEXT:    s_or_b32 s12, s73, s12
; VI-NEXT:    s_or_b32 s13, s72, s13
; VI-NEXT:    s_or_b32 s14, s63, s14
; VI-NEXT:    s_or_b32 s15, s62, s15
; VI-NEXT:    s_or_b32 s40, s61, s40
; VI-NEXT:    s_or_b32 s41, s60, s41
; VI-NEXT:    s_or_b32 s42, s59, s42
; VI-NEXT:    s_or_b32 s43, s58, s43
; VI-NEXT:    s_or_b32 s44, s57, s44
; VI-NEXT:    s_or_b32 s45, s56, s45
; VI-NEXT:    s_or_b32 s46, s47, s46
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s6, s6, 0x30000
; VI-NEXT:    s_add_i32 s7, s7, 0x30000
; VI-NEXT:    s_add_i32 s29, s29, 0x30000
; VI-NEXT:    s_add_i32 s28, s28, 0x30000
; VI-NEXT:    s_add_i32 s27, s27, 0x30000
; VI-NEXT:    s_add_i32 s26, s26, 0x30000
; VI-NEXT:    s_add_i32 s25, s25, 0x30000
; VI-NEXT:    s_add_i32 s24, s24, 0x30000
; VI-NEXT:    s_add_i32 s23, s23, 0x30000
; VI-NEXT:    s_add_i32 s22, s22, 0x30000
; VI-NEXT:    s_add_i32 s21, s21, 0x30000
; VI-NEXT:    s_add_i32 s20, s20, 0x30000
; VI-NEXT:    s_add_i32 s19, s19, 0x30000
; VI-NEXT:    s_add_i32 s18, s18, 0x30000
; VI-NEXT:    s_add_i32 s17, s17, 0x30000
; VI-NEXT:    s_add_i32 s16, s16, 0x30000
; VI-NEXT:    s_add_i32 s8, s8, 0x30000
; VI-NEXT:    s_add_i32 s9, s9, 0x30000
; VI-NEXT:    s_add_i32 s10, s10, 0x30000
; VI-NEXT:    s_add_i32 s11, s11, 0x30000
; VI-NEXT:    s_add_i32 s12, s12, 0x30000
; VI-NEXT:    s_add_i32 s13, s13, 0x30000
; VI-NEXT:    s_add_i32 s14, s14, 0x30000
; VI-NEXT:    s_add_i32 s15, s15, 0x30000
; VI-NEXT:    s_add_i32 s40, s40, 0x30000
; VI-NEXT:    s_add_i32 s41, s41, 0x30000
; VI-NEXT:    s_add_i32 s42, s42, 0x30000
; VI-NEXT:    s_add_i32 s43, s43, 0x30000
; VI-NEXT:    s_add_i32 s44, s44, 0x30000
; VI-NEXT:    s_add_i32 s45, s45, 0x30000
; VI-NEXT:    s_add_i32 s46, s46, 0x30000
; VI-NEXT:    s_add_i32 s47, s4, 0x30000
; VI-NEXT:  .LBB107_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s7
; VI-NEXT:    v_mov_b32_e32 v15, s6
; VI-NEXT:    v_mov_b32_e32 v16, s47
; VI-NEXT:    v_mov_b32_e32 v17, s46
; VI-NEXT:    v_mov_b32_e32 v18, s45
; VI-NEXT:    v_mov_b32_e32 v19, s44
; VI-NEXT:    v_mov_b32_e32 v20, s43
; VI-NEXT:    v_mov_b32_e32 v21, s42
; VI-NEXT:    v_mov_b32_e32 v22, s41
; VI-NEXT:    v_mov_b32_e32 v23, s40
; VI-NEXT:    v_mov_b32_e32 v24, s15
; VI-NEXT:    v_mov_b32_e32 v25, s14
; VI-NEXT:    v_mov_b32_e32 v26, s13
; VI-NEXT:    v_mov_b32_e32 v27, s12
; VI-NEXT:    v_mov_b32_e32 v28, s11
; VI-NEXT:    v_mov_b32_e32 v29, s10
; VI-NEXT:    v_mov_b32_e32 v30, s9
; VI-NEXT:    v_mov_b32_e32 v31, s8
; VI-NEXT:    v_readlane_b32 s39, v32, 7
; VI-NEXT:    v_readlane_b32 s38, v32, 6
; VI-NEXT:    v_readlane_b32 s37, v32, 5
; VI-NEXT:    v_readlane_b32 s36, v32, 4
; VI-NEXT:    v_readlane_b32 s35, v32, 3
; VI-NEXT:    v_readlane_b32 s34, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB107_4:
; VI-NEXT:    s_branch .LBB107_2
;
; GFX9-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB107_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB107_4
; GFX9-NEXT:  .LBB107_2: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, s31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, s30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, s29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, s28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, s27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, s26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, s25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, s24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, s23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, s22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, s21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, s20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, s19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, s18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, s17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, s16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v31, s51, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, s50, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, s49, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, s48, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, s47, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, s46, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, s45, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, s44, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, s43, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, s42, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, s41, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, s40, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, s39, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, s38, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, s37, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, s36, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB107_5
; GFX9-NEXT:  .LBB107_3:
; GFX9-NEXT:    s_branch .LBB107_2
; GFX9-NEXT:  .LBB107_4:
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:  .LBB107_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v64bf16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB107_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB107_4
; GFX11-NEXT:  .LBB107_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, s24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, s23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, s22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, s21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, s20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, s19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, s18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, s17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, s16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, s15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, s14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, s13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, s12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v31, s51, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, s50, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, s49, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, s48, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, s47, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, s46, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, s45, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, s44, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, s43, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, s42, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, s41, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, s40, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, s39, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, s38, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, s37, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, s36, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_branch .LBB107_5
; GFX11-NEXT:  .LBB107_3:
; GFX11-NEXT:    s_branch .LBB107_2
; GFX11-NEXT:  .LBB107_4:
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:  .LBB107_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <64 x bfloat>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <64 x bfloat>
  br label %end

end:
  %phi = phi <64 x bfloat> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x bfloat> %phi
}

define <64 x i16> @bitcast_v64f16_to_v64i16(<64 x half> %a, i32 %b) {
; SI-LABEL: bitcast_v64f16_to_v64i16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v30
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v28
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v24
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v22
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v20
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v45, 16, v18
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v17
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v16
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v15
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v14
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v13
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v12
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v11
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v10
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v9
; SI-NEXT:    s_waitcnt expcnt(4)
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v8
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v7
; SI-NEXT:    s_waitcnt expcnt(3)
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v5
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshrrev_b32_e32 v61, 16, v4
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v3
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v62, 16, v2
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v1
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v0
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v37
; SI-NEXT:    v_lshrrev_b32_e32 v37, 16, v31
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB108_2
; SI-NEXT:  ; %bb.1: ; %cmp.true
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v32, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v32, v32
; SI-NEXT:    buffer_store_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v30, v36
; SI-NEXT:    v_cvt_f32_f16_e32 v37, v37
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v36, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f32_f16_e32 v31, v31
; SI-NEXT:    v_cvt_f32_f16_e32 v29, v29
; SI-NEXT:    v_cvt_f32_f16_e32 v33, v33
; SI-NEXT:    v_cvt_f32_f16_e32 v27, v27
; SI-NEXT:    v_cvt_f32_f16_e32 v34, v34
; SI-NEXT:    v_add_f32_e32 v37, 0x38000000, v37
; SI-NEXT:    v_cvt_f32_f16_e32 v25, v25
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v37, v37
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    v_add_f32_e32 v27, 0x38000000, v27
; SI-NEXT:    v_add_f32_e32 v34, 0x38000000, v34
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v27
; SI-NEXT:    v_cvt_f16_f32_e32 v34, v34
; SI-NEXT:    v_add_f32_e32 v25, 0x38000000, v25
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v37
; SI-NEXT:    v_cvt_f16_f32_e32 v25, v25
; SI-NEXT:    v_or_b32_e32 v31, v31, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v36
; SI-NEXT:    v_or_b32_e32 v29, v29, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v33
; SI-NEXT:    v_or_b32_e32 v27, v27, v35
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v34
; SI-NEXT:    v_or_b32_e32 v25, v25, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v23, v23
; SI-NEXT:    v_cvt_f32_f16_e32 v21, v21
; SI-NEXT:    v_cvt_f32_f16_e32 v19, v19
; SI-NEXT:    v_cvt_f32_f16_e32 v17, v17
; SI-NEXT:    v_add_f32_e32 v23, 0x38000000, v23
; SI-NEXT:    v_cvt_f16_f32_e32 v23, v23
; SI-NEXT:    v_add_f32_e32 v21, 0x38000000, v21
; SI-NEXT:    v_cvt_f16_f32_e32 v21, v21
; SI-NEXT:    v_add_f32_e32 v19, 0x38000000, v19
; SI-NEXT:    v_cvt_f16_f32_e32 v19, v19
; SI-NEXT:    v_add_f32_e32 v17, 0x38000000, v17
; SI-NEXT:    v_cvt_f16_f32_e32 v17, v17
; SI-NEXT:    v_cvt_f32_f16_e32 v15, v15
; SI-NEXT:    v_cvt_f32_f16_e32 v13, v13
; SI-NEXT:    v_cvt_f32_f16_e32 v11, v11
; SI-NEXT:    v_cvt_f32_f16_e32 v9, v9
; SI-NEXT:    v_add_f32_e32 v15, 0x38000000, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v15, v15
; SI-NEXT:    v_add_f32_e32 v13, 0x38000000, v13
; SI-NEXT:    v_cvt_f16_f32_e32 v13, v13
; SI-NEXT:    v_add_f32_e32 v11, 0x38000000, v11
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v11
; SI-NEXT:    v_add_f32_e32 v9, 0x38000000, v9
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v9
; SI-NEXT:    v_cvt_f32_f16_e32 v7, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v5, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v3, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v7, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v7
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v62, v62
; SI-NEXT:    v_cvt_f32_f16_e32 v61, v61
; SI-NEXT:    v_cvt_f32_f16_e32 v60, v60
; SI-NEXT:    v_cvt_f32_f16_e32 v59, v59
; SI-NEXT:    v_cvt_f32_f16_e32 v58, v58
; SI-NEXT:    v_cvt_f32_f16_e32 v57, v57
; SI-NEXT:    v_cvt_f32_f16_e32 v56, v56
; SI-NEXT:    v_cvt_f32_f16_e32 v47, v47
; SI-NEXT:    v_cvt_f32_f16_e32 v45, v45
; SI-NEXT:    v_cvt_f32_f16_e32 v42, v42
; SI-NEXT:    v_cvt_f32_f16_e32 v55, v55
; SI-NEXT:    v_cvt_f32_f16_e32 v52, v52
; SI-NEXT:    v_cvt_f32_f16_e32 v49, v49
; SI-NEXT:    v_cvt_f32_f16_e32 v38, v38
; SI-NEXT:    v_cvt_f32_f16_e32 v0, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v2, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v4, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v6, v6
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v8, v8
; SI-NEXT:    v_cvt_f32_f16_e32 v10, v10
; SI-NEXT:    v_cvt_f32_f16_e32 v12, v12
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v32
; SI-NEXT:    v_or_b32_e32 v23, v23, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v39
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    v_cvt_f32_f16_e32 v14, v14
; SI-NEXT:    v_cvt_f32_f16_e32 v16, v16
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v18, v18
; SI-NEXT:    v_cvt_f32_f16_e32 v20, v20
; SI-NEXT:    v_cvt_f32_f16_e32 v22, v22
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v32
; SI-NEXT:    v_or_b32_e32 v21, v21, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v48
; SI-NEXT:    v_cvt_f32_f16_e32 v24, v24
; SI-NEXT:    v_cvt_f32_f16_e32 v26, v26
; SI-NEXT:    v_cvt_f32_f16_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v39, v35
; SI-NEXT:    v_add_f32_e32 v62, 0x38000000, v62
; SI-NEXT:    v_add_f32_e32 v61, 0x38000000, v61
; SI-NEXT:    v_add_f32_e32 v60, 0x38000000, v60
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v39
; SI-NEXT:    v_or_b32_e32 v19, v19, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v50
; SI-NEXT:    v_add_f32_e32 v59, 0x38000000, v59
; SI-NEXT:    v_add_f32_e32 v58, 0x38000000, v58
; SI-NEXT:    v_add_f32_e32 v57, 0x38000000, v57
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v48, v35
; SI-NEXT:    v_add_f32_e32 v56, 0x38000000, v56
; SI-NEXT:    v_add_f32_e32 v47, 0x38000000, v47
; SI-NEXT:    v_add_f32_e32 v45, 0x38000000, v45
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v48
; SI-NEXT:    v_or_b32_e32 v17, v17, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v51
; SI-NEXT:    v_add_f32_e32 v42, 0x38000000, v42
; SI-NEXT:    v_add_f32_e32 v55, 0x38000000, v55
; SI-NEXT:    v_add_f32_e32 v52, 0x38000000, v52
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v50, v35
; SI-NEXT:    v_add_f32_e32 v49, 0x38000000, v49
; SI-NEXT:    v_add_f32_e32 v38, 0x38000000, v38
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v50
; SI-NEXT:    v_or_b32_e32 v15, v15, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v53
; SI-NEXT:    v_cvt_f16_f32_e32 v62, v62
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v61, v61
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v51, v35
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v60, v60
; SI-NEXT:    v_add_f32_e32 v6, 0x38000000, v6
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v51
; SI-NEXT:    v_or_b32_e32 v13, v13, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v54
; SI-NEXT:    v_cvt_f16_f32_e32 v59, v59
; SI-NEXT:    v_add_f32_e32 v8, 0x38000000, v8
; SI-NEXT:    v_cvt_f16_f32_e32 v58, v58
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v53, v35
; SI-NEXT:    v_add_f32_e32 v10, 0x38000000, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v57, v57
; SI-NEXT:    v_add_f32_e32 v12, 0x38000000, v12
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v53
; SI-NEXT:    v_or_b32_e32 v11, v11, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v40
; SI-NEXT:    v_cvt_f16_f32_e32 v56, v56
; SI-NEXT:    v_add_f32_e32 v14, 0x38000000, v14
; SI-NEXT:    v_cvt_f16_f32_e32 v47, v47
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v54, v35
; SI-NEXT:    v_add_f32_e32 v16, 0x38000000, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v45, v45
; SI-NEXT:    v_add_f32_e32 v18, 0x38000000, v18
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v54
; SI-NEXT:    v_or_b32_e32 v9, v9, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v41
; SI-NEXT:    v_cvt_f16_f32_e32 v42, v42
; SI-NEXT:    v_add_f32_e32 v20, 0x38000000, v20
; SI-NEXT:    v_cvt_f16_f32_e32 v55, v55
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v40, v35
; SI-NEXT:    v_add_f32_e32 v22, 0x38000000, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v52, v52
; SI-NEXT:    v_add_f32_e32 v24, 0x38000000, v24
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v40
; SI-NEXT:    v_or_b32_e32 v7, v7, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v43
; SI-NEXT:    v_cvt_f16_f32_e32 v49, v49
; SI-NEXT:    v_add_f32_e32 v26, 0x38000000, v26
; SI-NEXT:    v_cvt_f16_f32_e32 v38, v38
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v43, v35
; SI-NEXT:    v_add_f32_e32 v28, 0x38000000, v28
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v43
; SI-NEXT:    v_or_b32_e32 v5, v5, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v44
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v8, v8
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v44, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v10, v10
; SI-NEXT:    v_cvt_f16_f32_e32 v12, v12
; SI-NEXT:    v_cvt_f16_f32_e32 v14, v14
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v44
; SI-NEXT:    v_or_b32_e32 v3, v3, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v46
; SI-NEXT:    v_cvt_f16_f32_e32 v16, v16
; SI-NEXT:    v_cvt_f16_f32_e32 v18, v18
; SI-NEXT:    v_cvt_f16_f32_e32 v20, v20
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v46, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v22, v22
; SI-NEXT:    v_cvt_f16_f32_e32 v24, v24
; SI-NEXT:    v_cvt_f16_f32_e32 v26, v26
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v46
; SI-NEXT:    v_or_b32_e32 v1, v1, v35
; SI-NEXT:    v_cvt_f32_f16_e32 v35, v63
; SI-NEXT:    v_cvt_f32_f16_e32 v63, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v28
; SI-NEXT:    v_add_f32_e32 v35, 0x38000000, v35
; SI-NEXT:    v_add_f32_e32 v63, 0x38000000, v63
; SI-NEXT:    v_cvt_f16_f32_e32 v35, v35
; SI-NEXT:    v_cvt_f16_f32_e32 v63, v63
; SI-NEXT:    v_lshlrev_b32_e32 v62, 16, v62
; SI-NEXT:    v_lshlrev_b32_e32 v61, 16, v61
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v60
; SI-NEXT:    v_lshlrev_b32_e32 v59, 16, v59
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v58
; SI-NEXT:    v_lshlrev_b32_e32 v57, 16, v57
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v56
; SI-NEXT:    v_lshlrev_b32_e32 v47, 16, v47
; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v45
; SI-NEXT:    v_lshlrev_b32_e32 v42, 16, v42
; SI-NEXT:    v_lshlrev_b32_e32 v55, 16, v55
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v52
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v49
; SI-NEXT:    v_mov_b32_e32 v41, v40
; SI-NEXT:    v_mov_b32_e32 v40, v54
; SI-NEXT:    v_mov_b32_e32 v54, v53
; SI-NEXT:    v_mov_b32_e32 v53, v51
; SI-NEXT:    v_mov_b32_e32 v51, v50
; SI-NEXT:    v_mov_b32_e32 v50, v48
; SI-NEXT:    v_mov_b32_e32 v48, v39
; SI-NEXT:    v_mov_b32_e32 v39, v32
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v38
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v63
; SI-NEXT:    v_or_b32_e32 v0, v0, v35
; SI-NEXT:    v_or_b32_e32 v2, v2, v62
; SI-NEXT:    v_or_b32_e32 v4, v4, v61
; SI-NEXT:    v_or_b32_e32 v6, v6, v60
; SI-NEXT:    v_or_b32_e32 v8, v8, v59
; SI-NEXT:    v_or_b32_e32 v10, v10, v58
; SI-NEXT:    v_or_b32_e32 v12, v12, v57
; SI-NEXT:    v_or_b32_e32 v14, v14, v56
; SI-NEXT:    v_or_b32_e32 v16, v16, v47
; SI-NEXT:    v_or_b32_e32 v18, v18, v45
; SI-NEXT:    v_or_b32_e32 v20, v20, v42
; SI-NEXT:    v_or_b32_e32 v22, v22, v55
; SI-NEXT:    v_or_b32_e32 v24, v24, v52
; SI-NEXT:    v_or_b32_e32 v26, v26, v49
; SI-NEXT:    v_or_b32_e32 v28, v28, v38
; SI-NEXT:    v_alignbit_b32 v63, v1, v35, 16
; SI-NEXT:    v_mov_b32_e32 v35, v36
; SI-NEXT:    v_alignbit_b32 v62, v3, v62, 16
; SI-NEXT:    v_alignbit_b32 v61, v5, v61, 16
; SI-NEXT:    v_alignbit_b32 v60, v7, v60, 16
; SI-NEXT:    v_alignbit_b32 v59, v9, v59, 16
; SI-NEXT:    v_alignbit_b32 v58, v11, v58, 16
; SI-NEXT:    v_alignbit_b32 v57, v13, v57, 16
; SI-NEXT:    v_alignbit_b32 v56, v15, v56, 16
; SI-NEXT:    v_alignbit_b32 v47, v17, v47, 16
; SI-NEXT:    v_alignbit_b32 v45, v19, v45, 16
; SI-NEXT:    v_alignbit_b32 v42, v21, v42, 16
; SI-NEXT:    v_alignbit_b32 v55, v23, v55, 16
; SI-NEXT:    v_alignbit_b32 v52, v25, v52, 16
; SI-NEXT:    v_alignbit_b32 v49, v27, v49, 16
; SI-NEXT:    v_alignbit_b32 v38, v29, v38, 16
; SI-NEXT:    v_alignbit_b32 v36, v31, v32, 16
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v30, v30
; SI-NEXT:    v_add_f32_e32 v30, 0x38000000, v30
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:  .LBB108_2: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v63
; SI-NEXT:    v_or_b32_e32 v0, v0, v32
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v46
; SI-NEXT:    v_or_b32_e32 v1, v1, v32
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v62
; SI-NEXT:    v_or_b32_e32 v2, v2, v32
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v44
; SI-NEXT:    v_or_b32_e32 v3, v3, v32
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v61
; SI-NEXT:    v_or_b32_e32 v4, v4, v32
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v5
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v43
; SI-NEXT:    v_or_b32_e32 v5, v5, v32
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v60
; SI-NEXT:    v_or_b32_e32 v6, v6, v32
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v41
; SI-NEXT:    v_or_b32_e32 v7, v7, v32
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v59
; SI-NEXT:    v_or_b32_e32 v8, v8, v32
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v40
; SI-NEXT:    v_or_b32_e32 v9, v9, v32
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v58
; SI-NEXT:    v_or_b32_e32 v10, v10, v32
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v11
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v54
; SI-NEXT:    v_or_b32_e32 v11, v11, v32
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v57
; SI-NEXT:    v_or_b32_e32 v12, v12, v32
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v13
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v53
; SI-NEXT:    v_or_b32_e32 v13, v13, v32
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v56
; SI-NEXT:    v_or_b32_e32 v14, v14, v32
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v51
; SI-NEXT:    v_or_b32_e32 v15, v15, v32
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v47
; SI-NEXT:    v_or_b32_e32 v16, v16, v32
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v17
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v50
; SI-NEXT:    v_or_b32_e32 v17, v17, v32
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v45
; SI-NEXT:    v_or_b32_e32 v18, v18, v32
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v48
; SI-NEXT:    v_or_b32_e32 v19, v19, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v42
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v32
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v39
; SI-NEXT:    v_or_b32_e32 v21, v21, v32
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v55
; SI-NEXT:    v_or_b32_e32 v22, v22, v32
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v23
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v25
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_or_b32_e32 v23, v23, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v52
; SI-NEXT:    v_or_b32_e32 v24, v24, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v34
; SI-NEXT:    v_or_b32_e32 v25, v25, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v49
; SI-NEXT:    v_or_b32_e32 v26, v26, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; SI-NEXT:    v_or_b32_e32 v27, v27, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v38
; SI-NEXT:    v_or_b32_e32 v28, v28, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v35
; SI-NEXT:    v_or_b32_e32 v29, v29, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v36
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v37
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v64i16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB108_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 0x200
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v15
; VI-NEXT:    v_add_f16_sdwa v15, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v15, v33, v15
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v14
; VI-NEXT:    v_add_f16_sdwa v14, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v14, v33, v14
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v13
; VI-NEXT:    v_add_f16_sdwa v13, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v13, v33, v13
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v12
; VI-NEXT:    v_add_f16_sdwa v12, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v12, v33, v12
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v11
; VI-NEXT:    v_add_f16_sdwa v11, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v11, v33, v11
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v10
; VI-NEXT:    v_add_f16_sdwa v10, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v10, v33, v10
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v9
; VI-NEXT:    v_add_f16_sdwa v9, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v9, v33, v9
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v8
; VI-NEXT:    v_add_f16_sdwa v8, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v8, v33, v8
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v7
; VI-NEXT:    v_add_f16_sdwa v7, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v7, v33, v7
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v6
; VI-NEXT:    v_add_f16_sdwa v6, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v6, v33, v6
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v5
; VI-NEXT:    v_add_f16_sdwa v5, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v5, v33, v5
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v4
; VI-NEXT:    v_add_f16_sdwa v4, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v4, v33, v4
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v3
; VI-NEXT:    v_add_f16_sdwa v3, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v3, v33, v3
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v2
; VI-NEXT:    v_add_f16_sdwa v2, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v2, v33, v2
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v1
; VI-NEXT:    v_add_f16_sdwa v1, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v1, v33, v1
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v0
; VI-NEXT:    v_add_f16_sdwa v0, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v0, v33, v0
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v31
; VI-NEXT:    v_add_f16_sdwa v31, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v31, v33, v31
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v30
; VI-NEXT:    v_add_f16_sdwa v30, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v30, v33, v30
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v29
; VI-NEXT:    v_add_f16_sdwa v29, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v29, v33, v29
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v28
; VI-NEXT:    v_add_f16_sdwa v28, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v28, v33, v28
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v27
; VI-NEXT:    v_add_f16_sdwa v27, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v27, v33, v27
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v26
; VI-NEXT:    v_add_f16_sdwa v26, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v26, v33, v26
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v25
; VI-NEXT:    v_add_f16_sdwa v25, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v25, v33, v25
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v24
; VI-NEXT:    v_add_f16_sdwa v24, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v24, v33, v24
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v23
; VI-NEXT:    v_add_f16_sdwa v23, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v23, v33, v23
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v22
; VI-NEXT:    v_add_f16_sdwa v22, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v22, v33, v22
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v21
; VI-NEXT:    v_add_f16_sdwa v21, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v33, v21
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v20
; VI-NEXT:    v_add_f16_sdwa v20, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v20, v33, v20
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v19
; VI-NEXT:    v_add_f16_sdwa v19, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v19, v33, v19
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v18
; VI-NEXT:    v_add_f16_sdwa v18, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v18, v33, v18
; VI-NEXT:    v_add_f16_e32 v33, 0x200, v17
; VI-NEXT:    v_add_f16_sdwa v17, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, 0x200, v16
; VI-NEXT:    v_or_b32_e32 v17, v33, v17
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB108_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v64i16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB108_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    s_movk_i32 s6, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, v15, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, v14, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, v13, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, v12, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, v11, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, v10, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, v9, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, v8, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, v7, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, v6, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, v5, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, v4, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, v3, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, v2, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, v1, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, v0, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_f16 v31, v31, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, v30, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, v29, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, v28, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, v27, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, v26, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, v25, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, v24, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, v23, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, v22, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, v21, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, v20, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, v19, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, v18, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, v17, s6 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, v16, s6 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB108_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v64i16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB108_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, v0 op_sel_hi:[0,1]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, v24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, v22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, v21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, v19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1]
; GFX11-NEXT:  .LBB108_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define inreg <64 x i16> @bitcast_v64f16_to_v64i16_scalar(<64 x half> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64f16_to_v64i16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v63, s30, 0
; SI-NEXT:    v_writelane_b32 v63, s31, 1
; SI-NEXT:    v_writelane_b32 v63, s34, 2
; SI-NEXT:    v_writelane_b32 v63, s35, 3
; SI-NEXT:    v_writelane_b32 v63, s36, 4
; SI-NEXT:    v_writelane_b32 v63, s37, 5
; SI-NEXT:    v_writelane_b32 v63, s38, 6
; SI-NEXT:    v_readfirstlane_b32 s10, v17
; SI-NEXT:    v_readfirstlane_b32 s6, v16
; SI-NEXT:    v_readfirstlane_b32 s11, v15
; SI-NEXT:    v_readfirstlane_b32 s7, v14
; SI-NEXT:    v_readfirstlane_b32 s59, v13
; SI-NEXT:    v_readfirstlane_b32 s8, v12
; SI-NEXT:    v_readfirstlane_b32 s57, v11
; SI-NEXT:    v_readfirstlane_b32 s9, v10
; SI-NEXT:    v_readfirstlane_b32 s56, v9
; SI-NEXT:    v_readfirstlane_b32 s12, v8
; SI-NEXT:    v_readfirstlane_b32 s47, v7
; SI-NEXT:    v_readfirstlane_b32 s90, v6
; SI-NEXT:    v_readfirstlane_b32 s46, v5
; SI-NEXT:    v_readfirstlane_b32 s92, v4
; SI-NEXT:    v_readfirstlane_b32 s45, v3
; SI-NEXT:    v_readfirstlane_b32 s30, v2
; SI-NEXT:    v_readfirstlane_b32 s44, v1
; SI-NEXT:    v_readfirstlane_b32 s35, v0
; SI-NEXT:    v_writelane_b32 v63, s39, 7
; SI-NEXT:    s_lshr_b32 s43, s29, 16
; SI-NEXT:    s_lshr_b32 s93, s28, 16
; SI-NEXT:    s_lshr_b32 s42, s27, 16
; SI-NEXT:    s_lshr_b32 s31, s26, 16
; SI-NEXT:    s_lshr_b32 s41, s25, 16
; SI-NEXT:    s_lshr_b32 s34, s24, 16
; SI-NEXT:    s_lshr_b32 s40, s23, 16
; SI-NEXT:    s_lshr_b32 s36, s22, 16
; SI-NEXT:    s_lshr_b32 s15, s21, 16
; SI-NEXT:    s_lshr_b32 s37, s20, 16
; SI-NEXT:    s_lshr_b32 s14, s19, 16
; SI-NEXT:    s_lshr_b32 s38, s18, 16
; SI-NEXT:    s_lshr_b32 s13, s17, 16
; SI-NEXT:    s_lshr_b32 s39, s16, 16
; SI-NEXT:    s_lshr_b32 s88, s10, 16
; SI-NEXT:    s_lshr_b32 s61, s6, 16
; SI-NEXT:    s_lshr_b32 s76, s11, 16
; SI-NEXT:    s_lshr_b32 s63, s7, 16
; SI-NEXT:    s_lshr_b32 s78, s59, 16
; SI-NEXT:    s_lshr_b32 s74, s8, 16
; SI-NEXT:    s_lshr_b32 s75, s57, 16
; SI-NEXT:    s_lshr_b32 s77, s9, 16
; SI-NEXT:    s_lshr_b32 s73, s56, 16
; SI-NEXT:    s_lshr_b32 s79, s12, 16
; SI-NEXT:    s_lshr_b32 s72, s47, 16
; SI-NEXT:    s_lshr_b32 s89, s90, 16
; SI-NEXT:    s_lshr_b32 s62, s46, 16
; SI-NEXT:    s_lshr_b32 s91, s92, 16
; SI-NEXT:    s_lshr_b32 s60, s45, 16
; SI-NEXT:    s_lshr_b32 s94, s30, 16
; SI-NEXT:    s_lshr_b32 s58, s44, 16
; SI-NEXT:    s_lshr_b32 s95, s35, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    s_cbranch_scc0 .LBB109_3
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    s_cbranch_execnz .LBB109_4
; SI-NEXT:  .LBB109_2: ; %cmp.true
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s39
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s16
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s38
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s18
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s37
; SI-NEXT:    v_or_b32_e32 v40, v1, v38
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v3
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    s_waitcnt expcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v60, 16, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s20
; SI-NEXT:    v_or_b32_e32 v2, v2, v60
; SI-NEXT:    v_lshlrev_b32_e32 v56, 16, v0
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s36
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s22
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_or_b32_e32 v8, v1, v56
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s34
; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v0
; SI-NEXT:    v_or_b32_e32 v2, v2, v45
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s24
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s31
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s26
; SI-NEXT:    v_or_b32_e32 v10, v0, v12
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s93
; SI-NEXT:    v_lshlrev_b32_e32 v43, 16, v2
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s28
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v43
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s95
; SI-NEXT:    v_lshlrev_b32_e32 v41, 16, v0
; SI-NEXT:    v_or_b32_e32 v17, v2, v41
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s94
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s35
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_lshlrev_b32_e32 v52, 16, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s30
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s91
; SI-NEXT:    v_or_b32_e32 v21, v0, v52
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s89
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s74
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v1
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s79
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s63
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s77
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v7
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s61
; SI-NEXT:    v_cvt_f16_f32_e32 v28, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s88
; SI-NEXT:    v_lshlrev_b32_e32 v58, 16, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s78
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v30, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v6, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s59
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s75
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s57
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v5
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
; SI-NEXT:    v_or_b32_e32 v59, v2, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s73
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v9
; SI-NEXT:    v_or_b32_e32 v23, v5, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s56
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s72
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v5
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v7
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s47
; SI-NEXT:    v_or_b32_e32 v16, v2, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s62
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s46
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v55, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v9
; SI-NEXT:    v_or_b32_e32 v37, v5, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s60
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v55
; SI-NEXT:    v_or_b32_e32 v19, v4, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s45
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s58
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v47, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v7, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v47
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s44
; SI-NEXT:    v_or_b32_e32 v50, v2, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s43
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v7
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s29
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v35, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_or_b32_e32 v53, v5, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s42
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v35
; SI-NEXT:    v_or_b32_e32 v42, v4, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s27
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s41
; SI-NEXT:    v_cvt_f16_f32_e32 v34, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v5
; SI-NEXT:    v_cvt_f16_f32_e32 v11, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v34
; SI-NEXT:    v_cvt_f32_f16_e32 v5, s25
; SI-NEXT:    v_or_b32_e32 v44, v2, v4
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s40
; SI-NEXT:    v_cvt_f32_f16_e32 v7, s23
; SI-NEXT:    v_add_f32_e32 v5, 0x38000000, v5
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v5, v5
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_cvt_f16_f32_e32 v9, v4
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v7
; SI-NEXT:    v_cvt_f16_f32_e32 v4, v4
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v11
; SI-NEXT:    v_or_b32_e32 v13, v5, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s15
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v9
; SI-NEXT:    v_or_b32_e32 v46, v4, v5
; SI-NEXT:    v_cvt_f32_f16_e32 v4, s21
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s14
; SI-NEXT:    v_cvt_f16_f32_e32 v51, v2
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v4
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v4, 0x38000000, v31
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v51
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s19
; SI-NEXT:    v_or_b32_e32 v57, v2, v31
; SI-NEXT:    v_cvt_f32_f16_e32 v2, s13
; SI-NEXT:    v_cvt_f32_f16_e32 v33, s17
; SI-NEXT:    v_cvt_f16_f32_e32 v27, v4
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_add_f32_e32 v2, 0x38000000, v2
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v2, v2
; SI-NEXT:    v_add_f32_e32 v33, 0x38000000, v33
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v33
; SI-NEXT:    v_cvt_f32_f16_e32 v29, s76
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v27
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s92
; SI-NEXT:    v_or_b32_e32 v61, v32, v31
; SI-NEXT:    v_cvt_f32_f16_e32 v31, s11
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v2
; SI-NEXT:    v_or_b32_e32 v39, v33, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s10
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v29
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v3
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s90
; SI-NEXT:    v_cvt_f16_f32_e32 v33, v29
; SI-NEXT:    v_add_f32_e32 v29, 0x38000000, v31
; SI-NEXT:    v_cvt_f16_f32_e32 v29, v29
; SI-NEXT:    v_add_f32_e32 v31, 0x38000000, v32
; SI-NEXT:    v_cvt_f16_f32_e32 v31, v31
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v33
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_or_b32_e32 v29, v29, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v6
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s12
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_or_b32_e32 v26, v0, v49
; SI-NEXT:    v_or_b32_e32 v48, v3, v36
; SI-NEXT:    v_cvt_f32_f16_e32 v0, s9
; SI-NEXT:    v_cvt_f32_f16_e32 v3, s7
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_or_b32_e32 v20, v1, v18
; SI-NEXT:    v_cvt_f32_f16_e32 v1, s8
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_add_f32_e32 v0, 0x38000000, v0
; SI-NEXT:    v_add_f32_e32 v3, 0x38000000, v3
; SI-NEXT:    v_cvt_f16_f32_e32 v0, v0
; SI-NEXT:    v_add_f32_e32 v1, 0x38000000, v1
; SI-NEXT:    v_cvt_f16_f32_e32 v3, v3
; SI-NEXT:    v_or_b32_e32 v24, v32, v15
; SI-NEXT:    v_cvt_f16_f32_e32 v1, v1
; SI-NEXT:    v_cvt_f32_f16_e32 v32, s6
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    v_or_b32_e32 v25, v0, v22
; SI-NEXT:    v_or_b32_e32 v0, v3, v28
; SI-NEXT:    v_add_f32_e32 v32, 0x38000000, v32
; SI-NEXT:    v_or_b32_e32 v54, v1, v58
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[0:1], v[38:39], 16
; SI-NEXT:    v_mov_b32_e32 v38, v61
; SI-NEXT:    v_lshr_b64 v[60:61], v[60:61], 16
; SI-NEXT:    v_cvt_f16_f32_e32 v32, v32
; SI-NEXT:    v_mov_b32_e32 v1, v40
; SI-NEXT:    v_mov_b32_e32 v40, v57
; SI-NEXT:    v_lshr_b64 v[61:62], v[56:57], 16
; SI-NEXT:    v_lshr_b64 v[56:57], v[45:46], 16
; SI-NEXT:    v_lshr_b64 v[4:5], v[52:53], 16
; SI-NEXT:    v_mov_b32_e32 v57, v44
; SI-NEXT:    v_lshr_b64 v[44:45], v[43:44], 16
; SI-NEXT:    v_mov_b32_e32 v45, v42
; SI-NEXT:    v_lshr_b64 v[42:43], v[41:42], 16
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v7, v6
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    v_mov_b32_e32 v6, v8
; SI-NEXT:    v_mov_b32_e32 v8, v13
; SI-NEXT:    v_lshr_b64 v[13:14], v[12:13], 16
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshr_b64 v[4:5], v[49:50], 16
; SI-NEXT:    v_mov_b32_e32 v52, v16
; SI-NEXT:    v_lshr_b64 v[15:16], v[15:16], 16
; SI-NEXT:    v_mov_b32_e32 v43, v23
; SI-NEXT:    v_lshr_b64 v[22:23], v[22:23], 16
; SI-NEXT:    v_or_b32_e32 v32, v32, v30
; SI-NEXT:    v_mov_b32_e32 v14, v17
; SI-NEXT:    v_mov_b32_e32 v17, v21
; SI-NEXT:    v_mov_b32_e32 v21, v37
; SI-NEXT:    v_lshr_b64 v[36:37], v[36:37], 16
; SI-NEXT:    v_mov_b32_e32 v3, v39
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v15, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v22, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v23, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v39, v32
; SI-NEXT:    v_mov_b32_e32 v37, v33
; SI-NEXT:    v_lshr_b64 v[32:33], v[30:31], 16
; SI-NEXT:    s_waitcnt expcnt(6)
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[4:5], v[18:19], 16
; SI-NEXT:    v_mov_b32_e32 v15, v46
; SI-NEXT:    v_mov_b32_e32 v22, v50
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    v_lshr_b64 v[49:50], v[28:29], 16
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    v_mov_b32_e32 v41, v55
; SI-NEXT:    v_mov_b32_e32 v55, v59
; SI-NEXT:    v_lshr_b64 v[58:59], v[58:59], 16
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v50, v7
; SI-NEXT:    s_branch .LBB109_5
; SI-NEXT:  .LBB109_3:
; SI-NEXT:    s_branch .LBB109_2
; SI-NEXT:  .LBB109_4:
; SI-NEXT:    v_mov_b32_e32 v6, s20
; SI-NEXT:    v_mov_b32_e32 v7, s25
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v8, s22
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v7, s95
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v7, s94
; SI-NEXT:    v_mov_b32_e32 v0, s7
; SI-NEXT:    v_mov_b32_e32 v50, s88
; SI-NEXT:    v_mov_b32_e32 v37, s76
; SI-NEXT:    v_mov_b32_e32 v30, s78
; SI-NEXT:    v_mov_b32_e32 v28, s75
; SI-NEXT:    v_mov_b32_e32 v46, s73
; SI-NEXT:    v_mov_b32_e32 v23, s72
; SI-NEXT:    v_mov_b32_e32 v41, s62
; SI-NEXT:    v_mov_b32_e32 v47, s60
; SI-NEXT:    v_mov_b32_e32 v33, s58
; SI-NEXT:    v_mov_b32_e32 v35, s43
; SI-NEXT:    v_mov_b32_e32 v34, s42
; SI-NEXT:    v_mov_b32_e32 v11, s41
; SI-NEXT:    v_mov_b32_e32 v9, s40
; SI-NEXT:    v_mov_b32_e32 v51, s15
; SI-NEXT:    v_mov_b32_e32 v27, s14
; SI-NEXT:    v_mov_b32_e32 v2, s13
; SI-NEXT:    v_mov_b32_e32 v39, s6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    v_mov_b32_e32 v54, s8
; SI-NEXT:    v_mov_b32_e32 v25, s9
; SI-NEXT:    v_mov_b32_e32 v24, s12
; SI-NEXT:    v_mov_b32_e32 v48, s90
; SI-NEXT:    v_mov_b32_e32 v20, s92
; SI-NEXT:    v_mov_b32_e32 v26, s30
; SI-NEXT:    v_mov_b32_e32 v17, s35
; SI-NEXT:    v_mov_b32_e32 v14, s28
; SI-NEXT:    v_mov_b32_e32 v12, s26
; SI-NEXT:    v_mov_b32_e32 v10, s24
; SI-NEXT:    v_mov_b32_e32 v18, s18
; SI-NEXT:    v_mov_b32_e32 v1, s16
; SI-NEXT:    v_mov_b32_e32 v3, s17
; SI-NEXT:    v_mov_b32_e32 v38, s19
; SI-NEXT:    v_mov_b32_e32 v40, s21
; SI-NEXT:    v_mov_b32_e32 v15, s23
; SI-NEXT:    v_mov_b32_e32 v57, s27
; SI-NEXT:    v_mov_b32_e32 v45, s29
; SI-NEXT:    v_mov_b32_e32 v53, s44
; SI-NEXT:    v_mov_b32_e32 v22, s45
; SI-NEXT:    v_mov_b32_e32 v19, s46
; SI-NEXT:    v_mov_b32_e32 v21, s47
; SI-NEXT:    v_mov_b32_e32 v52, s56
; SI-NEXT:    v_mov_b32_e32 v43, s57
; SI-NEXT:    v_mov_b32_e32 v55, s59
; SI-NEXT:    v_mov_b32_e32 v29, s11
; SI-NEXT:    v_mov_b32_e32 v31, s10
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_mov_b32_e32 v0, s39
; SI-NEXT:    v_mov_b32_e32 v60, s38
; SI-NEXT:    v_mov_b32_e32 v61, s37
; SI-NEXT:    v_mov_b32_e32 v56, s36
; SI-NEXT:    v_mov_b32_e32 v13, s34
; SI-NEXT:    v_mov_b32_e32 v44, s31
; SI-NEXT:    v_mov_b32_e32 v42, s93
; SI-NEXT:    v_mov_b32_e32 v36, s89
; SI-NEXT:    v_mov_b32_e32 v58, s74
; SI-NEXT:    v_mov_b32_e32 v49, s63
; SI-NEXT:    v_mov_b32_e32 v32, s61
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v7, s91
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v7, s79
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_mov_b32_e32 v7, s77
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:  .LBB109_5: ; %end
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v60
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v2, v3, v2
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v38
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v27
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v61
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v4, v6, v4
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v40
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v51
; SI-NEXT:    v_or_b32_e32 v5, v6, v5
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v56
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v6, v8, v6
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v15
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v9
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v13
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v8, v10, v8
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v45
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v17
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v26
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v19
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v21
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v55
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v31
; SI-NEXT:    v_readlane_b32 s39, v63, 7
; SI-NEXT:    v_readlane_b32 s38, v63, 6
; SI-NEXT:    v_readlane_b32 s37, v63, 5
; SI-NEXT:    v_readlane_b32 s36, v63, 4
; SI-NEXT:    v_readlane_b32 s35, v63, 3
; SI-NEXT:    v_readlane_b32 s34, v63, 2
; SI-NEXT:    v_readlane_b32 s31, v63, 1
; SI-NEXT:    v_readlane_b32 s30, v63, 0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v11
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v44
; SI-NEXT:    v_or_b32_e32 v10, v12, v10
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v57
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v34
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v42
; SI-NEXT:    v_or_b32_e32 v12, v14, v12
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v35
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v53
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_or_b32_e32 v14, v16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v33
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v36
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_or_b32_e32 v16, v18, v16
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v47
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v20
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v33
; SI-NEXT:    v_or_b32_e32 v18, v18, v20
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v41
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v48
; SI-NEXT:    v_or_b32_e32 v20, v20, v22
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v23
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v24
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v23
; SI-NEXT:    v_or_b32_e32 v22, v22, v24
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v52
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v46
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v25
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v25
; SI-NEXT:    v_or_b32_e32 v24, v24, v26
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v43
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v58
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v54
; SI-NEXT:    v_or_b32_e32 v26, v26, v28
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v30
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v49
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v30
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v37
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v39
; SI-NEXT:    v_or_b32_e32 v30, v30, v32
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v50
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_or_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64f16_to_v64i16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v33, s30, 0
; VI-NEXT:    v_writelane_b32 v33, s31, 1
; VI-NEXT:    v_writelane_b32 v33, s36, 2
; VI-NEXT:    v_writelane_b32 v33, s37, 3
; VI-NEXT:    v_writelane_b32 v33, s38, 4
; VI-NEXT:    v_writelane_b32 v33, s39, 5
; VI-NEXT:    v_writelane_b32 v33, s48, 6
; VI-NEXT:    v_writelane_b32 v33, s49, 7
; VI-NEXT:    v_writelane_b32 v33, s50, 8
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v33, s51, 9
; VI-NEXT:    v_readfirstlane_b32 s51, v17
; VI-NEXT:    v_readfirstlane_b32 s50, v16
; VI-NEXT:    v_readfirstlane_b32 s49, v15
; VI-NEXT:    v_readfirstlane_b32 s48, v14
; VI-NEXT:    v_readfirstlane_b32 s47, v13
; VI-NEXT:    v_readfirstlane_b32 s46, v12
; VI-NEXT:    v_readfirstlane_b32 s45, v11
; VI-NEXT:    v_readfirstlane_b32 s44, v10
; VI-NEXT:    v_readfirstlane_b32 s43, v9
; VI-NEXT:    v_readfirstlane_b32 s42, v8
; VI-NEXT:    v_readfirstlane_b32 s41, v7
; VI-NEXT:    v_readfirstlane_b32 s40, v6
; VI-NEXT:    v_readfirstlane_b32 s39, v5
; VI-NEXT:    v_readfirstlane_b32 s38, v4
; VI-NEXT:    v_readfirstlane_b32 s37, v3
; VI-NEXT:    v_readfirstlane_b32 s36, v2
; VI-NEXT:    v_readfirstlane_b32 s31, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s30, v0
; VI-NEXT:    s_cbranch_scc0 .LBB109_3
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB109_4
; VI-NEXT:  .LBB109_2: ; %cmp.true
; VI-NEXT:    s_lshr_b32 s4, s50, 16
; VI-NEXT:    v_mov_b32_e32 v0, s4
; VI-NEXT:    s_lshr_b32 s4, s30, 16
; VI-NEXT:    v_mov_b32_e32 v16, 0x200
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    s_lshr_b32 s4, s31, 16
; VI-NEXT:    v_add_f16_sdwa v18, v0, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, s30, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_mov_b32_e32 v3, s4
; VI-NEXT:    s_lshr_b32 s4, s29, 16
; VI-NEXT:    v_or_b32_e32 v14, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s29, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s28, 16
; VI-NEXT:    v_or_b32_e32 v13, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s28, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s27, 16
; VI-NEXT:    v_or_b32_e32 v12, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s27, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s26, 16
; VI-NEXT:    v_or_b32_e32 v11, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s26, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s25, 16
; VI-NEXT:    v_or_b32_e32 v10, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s25, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s24, 16
; VI-NEXT:    v_or_b32_e32 v9, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s24, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s23, 16
; VI-NEXT:    v_or_b32_e32 v8, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s23, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s22, 16
; VI-NEXT:    v_or_b32_e32 v7, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s22, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s21, 16
; VI-NEXT:    v_or_b32_e32 v6, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s21, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s20, 16
; VI-NEXT:    v_or_b32_e32 v5, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s20, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s19, 16
; VI-NEXT:    v_or_b32_e32 v4, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v2, s31, v16
; VI-NEXT:    v_add_f16_sdwa v3, v3, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v0, s19, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s18, 16
; VI-NEXT:    v_or_b32_e32 v15, v2, v3
; VI-NEXT:    v_or_b32_e32 v3, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    v_add_f16_e32 v0, s18, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    s_lshr_b32 s4, s17, 16
; VI-NEXT:    v_or_b32_e32 v2, v0, v1
; VI-NEXT:    v_mov_b32_e32 v1, s4
; VI-NEXT:    s_lshr_b32 s4, s16, 16
; VI-NEXT:    v_mov_b32_e32 v19, s4
; VI-NEXT:    s_lshr_b32 s4, s51, 16
; VI-NEXT:    v_add_f16_e32 v17, s50, v16
; VI-NEXT:    v_mov_b32_e32 v20, s4
; VI-NEXT:    s_lshr_b32 s4, s49, 16
; VI-NEXT:    v_or_b32_e32 v30, v17, v18
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s49, v16
; VI-NEXT:    s_lshr_b32 s4, s48, 16
; VI-NEXT:    v_or_b32_e32 v29, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s48, v16
; VI-NEXT:    s_lshr_b32 s4, s47, 16
; VI-NEXT:    v_or_b32_e32 v28, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s47, v16
; VI-NEXT:    s_lshr_b32 s4, s46, 16
; VI-NEXT:    v_or_b32_e32 v27, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s46, v16
; VI-NEXT:    s_lshr_b32 s4, s45, 16
; VI-NEXT:    v_or_b32_e32 v26, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s45, v16
; VI-NEXT:    s_lshr_b32 s4, s44, 16
; VI-NEXT:    v_or_b32_e32 v25, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s44, v16
; VI-NEXT:    s_lshr_b32 s4, s43, 16
; VI-NEXT:    v_or_b32_e32 v24, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s43, v16
; VI-NEXT:    s_lshr_b32 s4, s42, 16
; VI-NEXT:    v_or_b32_e32 v23, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s42, v16
; VI-NEXT:    s_lshr_b32 s4, s41, 16
; VI-NEXT:    v_or_b32_e32 v22, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_e32 v0, s17, v16
; VI-NEXT:    v_add_f16_sdwa v1, v1, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s41, v16
; VI-NEXT:    s_lshr_b32 s4, s40, 16
; VI-NEXT:    v_or_b32_e32 v1, v0, v1
; VI-NEXT:    v_add_f16_e32 v0, s16, v16
; VI-NEXT:    v_add_f16_sdwa v19, v19, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_or_b32_e32 v21, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_or_b32_e32 v0, v0, v19
; VI-NEXT:    v_add_f16_e32 v19, s51, v16
; VI-NEXT:    v_add_f16_sdwa v20, v20, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s40, v16
; VI-NEXT:    s_lshr_b32 s4, s39, 16
; VI-NEXT:    v_or_b32_e32 v31, v19, v20
; VI-NEXT:    v_or_b32_e32 v20, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s39, v16
; VI-NEXT:    s_lshr_b32 s4, s38, 16
; VI-NEXT:    v_or_b32_e32 v19, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v18, s38, v16
; VI-NEXT:    s_lshr_b32 s4, s37, 16
; VI-NEXT:    v_or_b32_e32 v18, v18, v17
; VI-NEXT:    v_mov_b32_e32 v17, s4
; VI-NEXT:    v_add_f16_sdwa v17, v17, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v32, s37, v16
; VI-NEXT:    s_lshr_b32 s4, s36, 16
; VI-NEXT:    v_or_b32_e32 v17, v32, v17
; VI-NEXT:    v_mov_b32_e32 v32, s4
; VI-NEXT:    v_add_f16_sdwa v32, v32, v16 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
; VI-NEXT:    v_add_f16_e32 v16, s36, v16
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:    s_branch .LBB109_5
; VI-NEXT:  .LBB109_3:
; VI-NEXT:    s_branch .LBB109_2
; VI-NEXT:  .LBB109_4:
; VI-NEXT:    v_mov_b32_e32 v16, s36
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v17, s37
; VI-NEXT:    v_mov_b32_e32 v18, s38
; VI-NEXT:    v_mov_b32_e32 v19, s39
; VI-NEXT:    v_mov_b32_e32 v20, s40
; VI-NEXT:    v_mov_b32_e32 v21, s41
; VI-NEXT:    v_mov_b32_e32 v22, s42
; VI-NEXT:    v_mov_b32_e32 v23, s43
; VI-NEXT:    v_mov_b32_e32 v24, s44
; VI-NEXT:    v_mov_b32_e32 v25, s45
; VI-NEXT:    v_mov_b32_e32 v26, s46
; VI-NEXT:    v_mov_b32_e32 v27, s47
; VI-NEXT:    v_mov_b32_e32 v28, s48
; VI-NEXT:    v_mov_b32_e32 v29, s49
; VI-NEXT:    v_mov_b32_e32 v30, s50
; VI-NEXT:    v_mov_b32_e32 v31, s51
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s30
; VI-NEXT:    v_mov_b32_e32 v15, s31
; VI-NEXT:  .LBB109_5: ; %end
; VI-NEXT:    v_readlane_b32 s51, v33, 9
; VI-NEXT:    v_readlane_b32 s50, v33, 8
; VI-NEXT:    v_readlane_b32 s49, v33, 7
; VI-NEXT:    v_readlane_b32 s48, v33, 6
; VI-NEXT:    v_readlane_b32 s39, v33, 5
; VI-NEXT:    v_readlane_b32 s38, v33, 4
; VI-NEXT:    v_readlane_b32 s37, v33, 3
; VI-NEXT:    v_readlane_b32 s36, v33, 2
; VI-NEXT:    v_readlane_b32 s31, v33, 1
; VI-NEXT:    v_readlane_b32 s30, v33, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64f16_to_v64i16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB109_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB109_4
; GFX9-NEXT:  .LBB109_2: ; %cmp.true
; GFX9-NEXT:    v_mov_b32_e32 v16, 0x200
; GFX9-NEXT:    v_pk_add_f16 v15, s31, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v14, s30, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v13, s29, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v12, s28, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v11, s27, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v10, s26, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v9, s25, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v8, s24, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v7, s23, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v6, s22, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v5, s21, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v4, s20, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v3, s19, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v2, s18, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v1, s17, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v0, s16, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v31, s51, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v30, s50, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v29, s49, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v28, s48, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v27, s47, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v26, s46, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v25, s45, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v24, s44, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v23, s43, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v22, s42, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v21, s41, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v20, s40, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v19, s39, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v18, s38, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v17, s37, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_f16 v16, s36, v16 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB109_5
; GFX9-NEXT:  .LBB109_3:
; GFX9-NEXT:    s_branch .LBB109_2
; GFX9-NEXT:  .LBB109_4:
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:  .LBB109_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64f16_to_v64i16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB109_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB109_4
; GFX11-NEXT:  .LBB109_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_f16 v15, 0x200, s27 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v14, 0x200, s26 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v13, 0x200, s25 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v12, 0x200, s24 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v11, 0x200, s23 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v10, 0x200, s22 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v9, 0x200, s21 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v8, 0x200, s20 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v7, 0x200, s19 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v6, 0x200, s18 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v5, 0x200, s17 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v4, 0x200, s16 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v3, 0x200, s15 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v2, 0x200, s14 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v1, 0x200, s13 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v0, 0x200, s12 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v31, 0x200, s51 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v30, 0x200, s50 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v29, 0x200, s49 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v28, 0x200, s48 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v27, 0x200, s47 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v26, 0x200, s46 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v25, 0x200, s45 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v24, 0x200, s44 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v23, 0x200, s43 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v22, 0x200, s42 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v21, 0x200, s41 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v20, 0x200, s40 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v19, 0x200, s39 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v18, 0x200, s38 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v17, 0x200, s37 op_sel_hi:[0,1]
; GFX11-NEXT:    v_pk_add_f16 v16, 0x200, s36 op_sel_hi:[0,1]
; GFX11-NEXT:    s_branch .LBB109_5
; GFX11-NEXT:  .LBB109_3:
; GFX11-NEXT:    s_branch .LBB109_2
; GFX11-NEXT:  .LBB109_4:
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:  .LBB109_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = fadd <64 x half> %a, splat (half 0xH0200)
  %a2 = bitcast <64 x half> %a1 to <64 x i16>
  br label %end

cmp.false:
  %a3 = bitcast <64 x half> %a to <64 x i16>
  br label %end

end:
  %phi = phi <64 x i16> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x i16> %phi
}

define <64 x half> @bitcast_v64i16_to_v64f16(<64 x i16> %a, i32 %b) {
; SI-LABEL: bitcast_v64i16_to_v64f16:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    buffer_load_dword v52, off, s[0:3], s32 offset:4
; SI-NEXT:    buffer_load_dword v37, off, s[0:3], s32
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
; SI-NEXT:    v_lshrrev_b32_e32 v50, 16, v2
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:264 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v49, 16, v4
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v48, 16, v6
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:268 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v49
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v39, 16, v8
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:272 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v48
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v38, 16, v10
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:276 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v39
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v36, 16, v12
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:280 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v38
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(5)
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v28
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:284 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v36
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:288 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v0, 16, v58
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v1
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v24
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v3
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v58
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshlrev_b32_e32 v38, 16, v60
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v5
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:304 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v63
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v7
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:308 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v60
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v9
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:312 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v40
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v11
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:316 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v54
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v56, 16, v13
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:320 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v51
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v47, 16, v15
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:324 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v56, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v56
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v46, 16, v17
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:292 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v47
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    v_lshrrev_b32_e32 v57, 16, v30
; SI-NEXT:    v_lshrrev_b32_e32 v53, 16, v29
; SI-NEXT:    v_lshrrev_b32_e32 v55, 16, v27
; SI-NEXT:    v_lshrrev_b32_e32 v59, 16, v26
; SI-NEXT:    v_lshrrev_b32_e32 v41, 16, v25
; SI-NEXT:    v_lshrrev_b32_e32 v42, 16, v23
; SI-NEXT:    v_lshrrev_b32_e32 v31, 16, v22
; SI-NEXT:    v_lshrrev_b32_e32 v43, 16, v21
; SI-NEXT:    v_lshrrev_b32_e32 v33, 16, v20
; SI-NEXT:    v_lshrrev_b32_e32 v44, 16, v19
; SI-NEXT:    v_lshrrev_b32_e32 v32, 16, v18
; SI-NEXT:    v_lshrrev_b32_e32 v35, 16, v16
; SI-NEXT:    v_lshrrev_b32_e32 v34, 16, v14
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v52
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v37
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:296 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v46
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v34, 16, v34
; SI-NEXT:    v_lshlrev_b32_e32 v35, 16, v35
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v32
; SI-NEXT:    v_lshlrev_b32_e32 v36, 16, v33
; SI-NEXT:    v_lshlrev_b32_e32 v45, 16, v31
; SI-NEXT:    v_lshlrev_b32_e32 v33, 16, v59
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v57
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:300 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    v_lshlrev_b32_e32 v48, 16, v44
; SI-NEXT:    v_lshlrev_b32_e32 v49, 16, v43
; SI-NEXT:    v_lshlrev_b32_e32 v39, 16, v42
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v41
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v55
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v53
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v52
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v55, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v53, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr61
; SI-NEXT:    ; implicit-def: $vgpr47
; SI-NEXT:    ; implicit-def: $vgpr62
; SI-NEXT:    ; implicit-def: $vgpr59
; SI-NEXT:    ; implicit-def: $vgpr57
; SI-NEXT:    ; implicit-def: $vgpr56
; SI-NEXT:    ; implicit-def: $vgpr46
; SI-NEXT:    ; implicit-def: $vgpr44
; SI-NEXT:    ; implicit-def: $vgpr43
; SI-NEXT:    ; implicit-def: $vgpr42
; SI-NEXT:    ; implicit-def: $vgpr41
; SI-NEXT:    ; implicit-def: $vgpr55
; SI-NEXT:    ; implicit-def: $vgpr53
; SI-NEXT:    ; kill: killed $vgpr50
; SI-NEXT:    ; implicit-def: $vgpr50
; SI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; SI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB110_2
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_lshlrev_b32_e32 v50, 16, v58
; SI-NEXT:    v_or_b32_e32 v61, v1, v50
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v3
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v63
; SI-NEXT:    v_or_b32_e32 v47, v1, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v60
; SI-NEXT:    v_or_b32_e32 v62, v1, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v7
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v40
; SI-NEXT:    v_or_b32_e32 v59, v1, v3
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v9
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v54
; SI-NEXT:    v_or_b32_e32 v57, v1, v3
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v51
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_alignbit_b32 v1, v61, v5, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v11
; SI-NEXT:    v_or_b32_e32 v56, v1, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_alignbit_b32 v1, v47, v7, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v13
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_or_b32_e32 v46, v1, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v62, v9, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v44, v1, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    v_alignbit_b32 v1, v59, v11, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_or_b32_e32 v43, v1, v3
; SI-NEXT:    v_alignbit_b32 v1, v57, v13, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v19
; SI-NEXT:    v_or_b32_e32 v42, v1, v48
; SI-NEXT:    v_alignbit_b32 v1, v56, v15, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v21
; SI-NEXT:    v_or_b32_e32 v41, v1, v49
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_alignbit_b32 v1, v46, v17, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v23
; SI-NEXT:    v_or_b32_e32 v55, v1, v39
; SI-NEXT:    v_alignbit_b32 v1, v44, v34, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v25
; SI-NEXT:    v_or_b32_e32 v53, v1, v4
; SI-NEXT:    v_alignbit_b32 v1, v43, v35, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v4, v1, v6
; SI-NEXT:    v_alignbit_b32 v1, v42, v32, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v3, v1, v8
; SI-NEXT:    v_alignbit_b32 v1, v41, v36, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v37
; SI-NEXT:    v_or_b32_e32 v50, v1, v31
; SI-NEXT:    v_alignbit_b32 v1, v55, v45, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr19
; SI-NEXT:    ; implicit-def: $vgpr21
; SI-NEXT:    ; implicit-def: $vgpr23
; SI-NEXT:    ; implicit-def: $vgpr25
; SI-NEXT:    ; implicit-def: $vgpr27
; SI-NEXT:    ; implicit-def: $vgpr29
; SI-NEXT:    ; implicit-def: $vgpr37
; SI-NEXT:    ; implicit-def: $vgpr48
; SI-NEXT:    ; implicit-def: $vgpr49
; SI-NEXT:    ; implicit-def: $vgpr39
; SI-NEXT:    ; implicit-def: $vgpr6
; SI-NEXT:    ; implicit-def: $vgpr8
; SI-NEXT:    ; implicit-def: $vgpr31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v5
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v53, v38, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr5
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v7
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v4, v33, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr7
; SI-NEXT:    ; implicit-def: $vgpr4
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v9
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v1, v3, v0, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr3
; SI-NEXT:    ; implicit-def: $vgpr9
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v11
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_alignbit_b32 v1, v50, v2, 16
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    ; implicit-def: $vgpr11
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v1
; SI-NEXT:    v_or_b32_e32 v1, v1, v13
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v1, v1, v15
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v1, v1, v17
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v1, v1, v34
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v1, v1, v35
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v1, v1, v32
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v1, v1, v36
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v1, v1, v45
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v1, v1, v38
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v1, v1, v33
; SI-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v0, v0, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr1
; SI-NEXT:    ; implicit-def: $vgpr10
; SI-NEXT:    ; implicit-def: $vgpr12
; SI-NEXT:    ; implicit-def: $vgpr13
; SI-NEXT:    ; implicit-def: $vgpr14
; SI-NEXT:    ; implicit-def: $vgpr15
; SI-NEXT:    ; implicit-def: $vgpr16
; SI-NEXT:    ; implicit-def: $vgpr17
; SI-NEXT:    ; implicit-def: $vgpr18
; SI-NEXT:    ; implicit-def: $vgpr20
; SI-NEXT:    ; implicit-def: $vgpr22
; SI-NEXT:    ; implicit-def: $vgpr24
; SI-NEXT:    ; implicit-def: $vgpr26
; SI-NEXT:    ; implicit-def: $vgpr28
; SI-NEXT:    ; implicit-def: $vgpr30
; SI-NEXT:    ; implicit-def: $vgpr34
; SI-NEXT:    ; kill: killed $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr35
; SI-NEXT:    ; implicit-def: $vgpr32
; SI-NEXT:    ; implicit-def: $vgpr36
; SI-NEXT:    ; implicit-def: $vgpr45
; SI-NEXT:    ; implicit-def: $vgpr38
; SI-NEXT:    ; implicit-def: $vgpr33
; SI-NEXT:    ; implicit-def: $vgpr0
; SI-NEXT:    ; implicit-def: $vgpr2
; SI-NEXT:  .LBB110_2: ; %Flow
; SI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; SI-NEXT:    s_cbranch_execz .LBB110_4
; SI-NEXT:  ; %bb.3: ; %cmp.true
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v28
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v0, v28
; SI-NEXT:    s_mov_b32 s6, 0x30000
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v30
; SI-NEXT:    v_add_i32_e32 v28, vcc, 3, v29
; SI-NEXT:    v_add_i32_e32 v29, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v26
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v2, v2, v30
; SI-NEXT:    v_or_b32_e32 v8, v8, v28
; SI-NEXT:    v_or_b32_e32 v0, v33, v0
; SI-NEXT:    v_add_i32_e32 v28, vcc, 0x30000, v2
; SI-NEXT:    v_add_i32_e32 v2, vcc, s6, v8
; SI-NEXT:    v_add_i32_e32 v8, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v27
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v6, v0
; SI-NEXT:    v_add_i32_e32 v6, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v24
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v38, v0
; SI-NEXT:    v_add_i32_e32 v24, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v25
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v4, v0
; SI-NEXT:    v_add_i32_e32 v53, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v22
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v45, v0
; SI-NEXT:    v_add_i32_e32 v4, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v23
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v39, v0
; SI-NEXT:    v_add_i32_e32 v55, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v20
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v36, v0
; SI-NEXT:    v_add_i32_e32 v20, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v21
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v49, v0
; SI-NEXT:    v_add_i32_e32 v41, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v18
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v32, v0
; SI-NEXT:    v_add_i32_e32 v18, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v19
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v48, v0
; SI-NEXT:    v_add_i32_e32 v42, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v16
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v35, v0
; SI-NEXT:    v_add_i32_e32 v16, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v17
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:300 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_add_i32_e32 v30, vcc, 3, v37
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v31, v30
; SI-NEXT:    v_add_i32_e32 v50, vcc, s6, v30
; SI-NEXT:    v_lshrrev_b32_e32 v52, 16, v50
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v17, v0
; SI-NEXT:    v_add_i32_e32 v43, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v14
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_or_b32_e32 v0, v34, v0
; SI-NEXT:    v_add_i32_e32 v14, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v15
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:296 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v15, v0
; SI-NEXT:    v_add_i32_e32 v44, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:288 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v12, v0
; SI-NEXT:    v_add_i32_e32 v12, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v13
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:292 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v13, v0
; SI-NEXT:    v_add_i32_e32 v46, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:284 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v10, v0
; SI-NEXT:    v_add_i32_e32 v10, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v11
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:324 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v11, v0
; SI-NEXT:    v_add_i32_e32 v56, vcc, s6, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:280 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v51, 16, v56
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v11, v0
; SI-NEXT:    v_add_i32_e32 v11, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v9
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:320 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v9, v0
; SI-NEXT:    v_add_i32_e32 v57, vcc, s6, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:276 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v54, 16, v57
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v9, v0
; SI-NEXT:    v_add_i32_e32 v9, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v7
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:316 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v7, v0
; SI-NEXT:    v_add_i32_e32 v59, vcc, s6, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:272 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v40, 16, v59
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v7, v0
; SI-NEXT:    v_add_i32_e32 v7, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v5
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:312 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v5, v0
; SI-NEXT:    v_add_i32_e32 v62, vcc, s6, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:268 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v60, 16, v62
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v5, v0
; SI-NEXT:    v_add_i32_e32 v5, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v3
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:308 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v3, v0
; SI-NEXT:    v_add_i32_e32 v47, vcc, s6, v0
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:264 ; 4-byte Folded Reload
; SI-NEXT:    v_lshrrev_b32_e32 v63, 16, v47
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v0
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_or_b32_e32 v0, v3, v0
; SI-NEXT:    v_add_i32_e32 v3, vcc, s6, v0
; SI-NEXT:    v_add_i32_e32 v0, vcc, 3, v1
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:304 ; 4-byte Folded Reload
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:252 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_or_b32_e32 v0, v1, v0
; SI-NEXT:    v_add_i32_e32 v61, vcc, s6, v0
; SI-NEXT:    v_alignbit_b32 v0, v61, v3, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:244 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v47, v5, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:236 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v62, v7, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:228 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v59, v9, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v11, off, s[0:3], s32 offset:220 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v57, v11, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:208 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v56, v10, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:192 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v46, v12, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:196 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v44, v14, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:204 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v16, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v43, v16, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:212 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v42, v18, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:216 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v41, v20, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:224 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v55, v4, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:232 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v53, v24, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:240 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v6, v8, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:248 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v29, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v2, v29, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:256 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v28, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_alignbit_b32 v0, v50, v28, 16
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:260 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v46
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v44
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v43
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v42
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v41
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v55
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v53
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v6
; SI-NEXT:    v_lshrrev_b32_e32 v58, 16, v61
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_lshrrev_b32_e32 v0, 16, v2
; SI-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
; SI-NEXT:  .LBB110_4: ; %end
; SI-NEXT:    s_or_b64 exec, exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:252 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:192 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v13, off, s[0:3], s32 offset:196 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v2, 16, v58
; SI-NEXT:    v_lshlrev_b32_e32 v4, 16, v63
; SI-NEXT:    v_lshlrev_b32_e32 v6, 16, v60
; SI-NEXT:    v_lshlrev_b32_e32 v8, 16, v40
; SI-NEXT:    v_lshlrev_b32_e32 v10, 16, v54
; SI-NEXT:    v_lshlrev_b32_e32 v12, 16, v51
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v32, 16, v52
; SI-NEXT:    buffer_load_dword v15, off, s[0:3], s32 offset:204 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v17, off, s[0:3], s32 offset:212 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v19, off, s[0:3], s32 offset:216 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v21, off, s[0:3], s32 offset:224 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v23, off, s[0:3], s32 offset:232 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v25, off, s[0:3], s32 offset:240 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:248 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:256 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:260 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v0, 0xffff, v0
; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
; SI-NEXT:    v_or_b32_e32 v0, v0, v1
; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v61
; SI-NEXT:    v_or_b32_e32 v1, v1, v2
; SI-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:244 ; 4-byte Folded Reload
; SI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
; SI-NEXT:    v_lshlrev_b32_e32 v5, 16, v5
; SI-NEXT:    v_lshlrev_b32_e32 v7, 16, v7
; SI-NEXT:    v_lshlrev_b32_e32 v9, 16, v9
; SI-NEXT:    v_lshlrev_b32_e32 v11, 16, v11
; SI-NEXT:    v_lshlrev_b32_e32 v13, 16, v13
; SI-NEXT:    v_lshlrev_b32_e32 v14, 16, v14
; SI-NEXT:    v_lshlrev_b32_e32 v16, 16, v16
; SI-NEXT:    v_lshlrev_b32_e32 v18, 16, v18
; SI-NEXT:    v_lshlrev_b32_e32 v20, 16, v20
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_lshlrev_b32_e32 v22, 16, v22
; SI-NEXT:    s_waitcnt vmcnt(13)
; SI-NEXT:    v_lshlrev_b32_e32 v24, 16, v24
; SI-NEXT:    s_waitcnt vmcnt(12)
; SI-NEXT:    v_lshlrev_b32_e32 v26, 16, v26
; SI-NEXT:    s_waitcnt vmcnt(11)
; SI-NEXT:    v_lshlrev_b32_e32 v28, 16, v28
; SI-NEXT:    s_waitcnt vmcnt(10)
; SI-NEXT:    v_lshlrev_b32_e32 v30, 16, v30
; SI-NEXT:    s_waitcnt vmcnt(9)
; SI-NEXT:    v_lshlrev_b32_e32 v15, 16, v15
; SI-NEXT:    s_waitcnt vmcnt(8)
; SI-NEXT:    v_lshlrev_b32_e32 v17, 16, v17
; SI-NEXT:    s_waitcnt vmcnt(7)
; SI-NEXT:    v_lshlrev_b32_e32 v19, 16, v19
; SI-NEXT:    s_waitcnt vmcnt(6)
; SI-NEXT:    v_lshlrev_b32_e32 v21, 16, v21
; SI-NEXT:    s_waitcnt vmcnt(5)
; SI-NEXT:    v_lshlrev_b32_e32 v23, 16, v23
; SI-NEXT:    s_waitcnt vmcnt(4)
; SI-NEXT:    v_lshlrev_b32_e32 v25, 16, v25
; SI-NEXT:    s_waitcnt vmcnt(3)
; SI-NEXT:    v_lshlrev_b32_e32 v27, 16, v27
; SI-NEXT:    s_waitcnt vmcnt(2)
; SI-NEXT:    v_lshlrev_b32_e32 v29, 16, v29
; SI-NEXT:    s_waitcnt vmcnt(1)
; SI-NEXT:    v_lshlrev_b32_e32 v31, 16, v31
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v2, 0xffff, v2
; SI-NEXT:    v_or_b32_e32 v2, v2, v3
; SI-NEXT:    v_and_b32_e32 v3, 0xffff, v47
; SI-NEXT:    v_or_b32_e32 v3, v3, v4
; SI-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:236 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v4, 0xffff, v4
; SI-NEXT:    v_or_b32_e32 v4, v4, v5
; SI-NEXT:    v_and_b32_e32 v5, 0xffff, v62
; SI-NEXT:    v_or_b32_e32 v5, v5, v6
; SI-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:228 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v6, 0xffff, v6
; SI-NEXT:    v_or_b32_e32 v6, v6, v7
; SI-NEXT:    v_and_b32_e32 v7, 0xffff, v59
; SI-NEXT:    v_or_b32_e32 v7, v7, v8
; SI-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:220 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v8, 0xffff, v8
; SI-NEXT:    v_or_b32_e32 v8, v8, v9
; SI-NEXT:    v_and_b32_e32 v9, 0xffff, v57
; SI-NEXT:    v_or_b32_e32 v9, v9, v10
; SI-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:208 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v10, 0xffff, v10
; SI-NEXT:    v_or_b32_e32 v10, v10, v11
; SI-NEXT:    v_and_b32_e32 v11, 0xffff, v56
; SI-NEXT:    v_or_b32_e32 v11, v11, v12
; SI-NEXT:    buffer_load_dword v12, off, s[0:3], s32 offset:200 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v12, 0xffff, v12
; SI-NEXT:    v_or_b32_e32 v12, v12, v13
; SI-NEXT:    v_and_b32_e32 v13, 0xffff, v46
; SI-NEXT:    v_or_b32_e32 v13, v13, v14
; SI-NEXT:    buffer_load_dword v14, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v14, 0xffff, v14
; SI-NEXT:    v_or_b32_e32 v14, v14, v15
; SI-NEXT:    v_and_b32_e32 v15, 0xffff, v44
; SI-NEXT:    v_or_b32_e32 v15, v15, v16
; SI-NEXT:    buffer_load_dword v16, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v16, 0xffff, v16
; SI-NEXT:    v_or_b32_e32 v16, v16, v17
; SI-NEXT:    v_and_b32_e32 v17, 0xffff, v43
; SI-NEXT:    v_or_b32_e32 v17, v17, v18
; SI-NEXT:    buffer_load_dword v18, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v18, 0xffff, v18
; SI-NEXT:    v_or_b32_e32 v18, v18, v19
; SI-NEXT:    v_and_b32_e32 v19, 0xffff, v42
; SI-NEXT:    v_or_b32_e32 v19, v19, v20
; SI-NEXT:    buffer_load_dword v20, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v20, 0xffff, v20
; SI-NEXT:    v_or_b32_e32 v20, v20, v21
; SI-NEXT:    v_and_b32_e32 v21, 0xffff, v41
; SI-NEXT:    v_or_b32_e32 v21, v21, v22
; SI-NEXT:    buffer_load_dword v22, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v63, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v62, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v61, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v60, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v59, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v58, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v57, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v56, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(14)
; SI-NEXT:    v_and_b32_e32 v22, 0xffff, v22
; SI-NEXT:    v_or_b32_e32 v22, v22, v23
; SI-NEXT:    v_and_b32_e32 v23, 0xffff, v55
; SI-NEXT:    v_or_b32_e32 v23, v23, v24
; SI-NEXT:    buffer_load_dword v24, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v24, 0xffff, v24
; SI-NEXT:    v_or_b32_e32 v24, v24, v25
; SI-NEXT:    v_and_b32_e32 v25, 0xffff, v53
; SI-NEXT:    v_or_b32_e32 v25, v25, v26
; SI-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v26, 0xffff, v26
; SI-NEXT:    v_or_b32_e32 v26, v26, v27
; SI-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v27, 0xffff, v27
; SI-NEXT:    v_or_b32_e32 v27, v27, v28
; SI-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v28, 0xffff, v28
; SI-NEXT:    v_or_b32_e32 v28, v28, v29
; SI-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v29, 0xffff, v29
; SI-NEXT:    v_or_b32_e32 v29, v29, v30
; SI-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    v_and_b32_e32 v30, 0xffff, v30
; SI-NEXT:    v_or_b32_e32 v30, v30, v31
; SI-NEXT:    v_and_b32_e32 v31, 0xffff, v50
; SI-NEXT:    v_or_b32_e32 v31, v31, v32
; SI-NEXT:    s_setpc_b64 s[30:31]
;
; VI-LABEL: bitcast_v64i16_to_v64f16:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:4
; VI-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; VI-NEXT:    s_waitcnt vmcnt(1)
; VI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v32
; VI-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; VI-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; VI-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; VI-NEXT:    s_cbranch_execz .LBB110_2
; VI-NEXT:  ; %bb.1: ; %cmp.true
; VI-NEXT:    v_mov_b32_e32 v32, 3
; VI-NEXT:    v_add_u16_sdwa v33, v15, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v15, 3, v15
; VI-NEXT:    v_or_b32_e32 v15, v15, v33
; VI-NEXT:    v_add_u16_sdwa v33, v14, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v14, 3, v14
; VI-NEXT:    v_or_b32_e32 v14, v14, v33
; VI-NEXT:    v_add_u16_sdwa v33, v13, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v13, 3, v13
; VI-NEXT:    v_or_b32_e32 v13, v13, v33
; VI-NEXT:    v_add_u16_sdwa v33, v12, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v12, 3, v12
; VI-NEXT:    v_or_b32_e32 v12, v12, v33
; VI-NEXT:    v_add_u16_sdwa v33, v11, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v11, 3, v11
; VI-NEXT:    v_or_b32_e32 v11, v11, v33
; VI-NEXT:    v_add_u16_sdwa v33, v10, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v10, 3, v10
; VI-NEXT:    v_or_b32_e32 v10, v10, v33
; VI-NEXT:    v_add_u16_sdwa v33, v9, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v9, 3, v9
; VI-NEXT:    v_or_b32_e32 v9, v9, v33
; VI-NEXT:    v_add_u16_sdwa v33, v8, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v8, 3, v8
; VI-NEXT:    v_or_b32_e32 v8, v8, v33
; VI-NEXT:    v_add_u16_sdwa v33, v7, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v7, 3, v7
; VI-NEXT:    v_or_b32_e32 v7, v7, v33
; VI-NEXT:    v_add_u16_sdwa v33, v6, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v6, 3, v6
; VI-NEXT:    v_or_b32_e32 v6, v6, v33
; VI-NEXT:    v_add_u16_sdwa v33, v5, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v5, 3, v5
; VI-NEXT:    v_or_b32_e32 v5, v5, v33
; VI-NEXT:    v_add_u16_sdwa v33, v4, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v4, 3, v4
; VI-NEXT:    v_or_b32_e32 v4, v4, v33
; VI-NEXT:    v_add_u16_sdwa v33, v3, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v3, 3, v3
; VI-NEXT:    v_or_b32_e32 v3, v3, v33
; VI-NEXT:    v_add_u16_sdwa v33, v2, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v2, 3, v2
; VI-NEXT:    v_or_b32_e32 v2, v2, v33
; VI-NEXT:    v_add_u16_sdwa v33, v1, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v1, 3, v1
; VI-NEXT:    v_or_b32_e32 v1, v1, v33
; VI-NEXT:    v_add_u16_sdwa v33, v0, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v0, 3, v0
; VI-NEXT:    v_or_b32_e32 v0, v0, v33
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    v_add_u16_sdwa v33, v31, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v31, 3, v31
; VI-NEXT:    v_or_b32_e32 v31, v31, v33
; VI-NEXT:    v_add_u16_sdwa v33, v30, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v30, 3, v30
; VI-NEXT:    v_or_b32_e32 v30, v30, v33
; VI-NEXT:    v_add_u16_sdwa v33, v29, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v29, 3, v29
; VI-NEXT:    v_or_b32_e32 v29, v29, v33
; VI-NEXT:    v_add_u16_sdwa v33, v28, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v28, 3, v28
; VI-NEXT:    v_or_b32_e32 v28, v28, v33
; VI-NEXT:    v_add_u16_sdwa v33, v27, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v27, 3, v27
; VI-NEXT:    v_or_b32_e32 v27, v27, v33
; VI-NEXT:    v_add_u16_sdwa v33, v26, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v26, 3, v26
; VI-NEXT:    v_or_b32_e32 v26, v26, v33
; VI-NEXT:    v_add_u16_sdwa v33, v25, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v25, 3, v25
; VI-NEXT:    v_or_b32_e32 v25, v25, v33
; VI-NEXT:    v_add_u16_sdwa v33, v24, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v24, 3, v24
; VI-NEXT:    v_or_b32_e32 v24, v24, v33
; VI-NEXT:    v_add_u16_sdwa v33, v23, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v23, 3, v23
; VI-NEXT:    v_or_b32_e32 v23, v23, v33
; VI-NEXT:    v_add_u16_sdwa v33, v22, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v22, 3, v22
; VI-NEXT:    v_or_b32_e32 v22, v22, v33
; VI-NEXT:    v_add_u16_sdwa v33, v21, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v21, 3, v21
; VI-NEXT:    v_or_b32_e32 v21, v21, v33
; VI-NEXT:    v_add_u16_sdwa v33, v20, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v20, 3, v20
; VI-NEXT:    v_or_b32_e32 v20, v20, v33
; VI-NEXT:    v_add_u16_sdwa v33, v19, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v19, 3, v19
; VI-NEXT:    v_or_b32_e32 v19, v19, v33
; VI-NEXT:    v_add_u16_sdwa v33, v18, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v18, 3, v18
; VI-NEXT:    v_or_b32_e32 v18, v18, v33
; VI-NEXT:    v_add_u16_sdwa v33, v17, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v17, 3, v17
; VI-NEXT:    v_add_u16_sdwa v32, v16, v32 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
; VI-NEXT:    v_add_u16_e32 v16, 3, v16
; VI-NEXT:    v_or_b32_e32 v17, v17, v33
; VI-NEXT:    v_or_b32_e32 v16, v16, v32
; VI-NEXT:  .LBB110_2: ; %end
; VI-NEXT:    s_or_b64 exec, exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
;
; GFX9-LABEL: bitcast_v64i16_to_v64f16:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:4
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v31
; GFX9-NEXT:    buffer_load_dword v31, off, s[0:3], s32
; GFX9-NEXT:    s_and_saveexec_b64 s[4:5], vcc
; GFX9-NEXT:    s_xor_b64 s[4:5], exec, s[4:5]
; GFX9-NEXT:    s_andn2_saveexec_b64 s[4:5], s[4:5]
; GFX9-NEXT:    s_cbranch_execz .LBB110_2
; GFX9-NEXT:  ; %bb.1: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:  .LBB110_2: ; %end
; GFX9-NEXT:    s_or_b64 exec, exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v64f16:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_clause 0x1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 offset:4
; GFX11-NEXT:    scratch_load_b32 v31, off, s32
; GFX11-NEXT:    s_mov_b32 s0, exec_lo
; GFX11-NEXT:    s_waitcnt vmcnt(1)
; GFX11-NEXT:    v_cmpx_ne_u32_e32 0, v32
; GFX11-NEXT:    s_xor_b32 s0, exec_lo, s0
; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
; GFX11-NEXT:    s_and_not1_saveexec_b32 s0, s0
; GFX11-NEXT:    s_cbranch_execz .LBB110_2
; GFX11-NEXT:  ; %bb.1: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, v0, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, v24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, v22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, v21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, v19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:  .LBB110_2: ; %end
; GFX11-NEXT:    s_or_b32 exec_lo, exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}

define inreg <64 x half> @bitcast_v64i16_to_v64f16_scalar(<64 x i16> inreg %a, i32 inreg %b) {
; SI-LABEL: bitcast_v64i16_to_v64f16_scalar:
; SI:       ; %bb.0:
; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; SI-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt expcnt(1)
; SI-NEXT:    v_writelane_b32 v32, s30, 0
; SI-NEXT:    v_writelane_b32 v32, s31, 1
; SI-NEXT:    v_writelane_b32 v32, s34, 2
; SI-NEXT:    v_writelane_b32 v32, s35, 3
; SI-NEXT:    v_writelane_b32 v32, s36, 4
; SI-NEXT:    v_writelane_b32 v32, s37, 5
; SI-NEXT:    v_writelane_b32 v32, s38, 6
; SI-NEXT:    v_writelane_b32 v32, s39, 7
; SI-NEXT:    v_writelane_b32 v32, s48, 8
; SI-NEXT:    v_writelane_b32 v32, s49, 9
; SI-NEXT:    v_writelane_b32 v32, s50, 10
; SI-NEXT:    v_writelane_b32 v32, s51, 11
; SI-NEXT:    s_lshr_b32 s4, s27, 16
; SI-NEXT:    ; implicit-def: $vgpr33 : SGPR spill to VGPR lane
; SI-NEXT:    v_writelane_b32 v32, s52, 12
; SI-NEXT:    s_waitcnt expcnt(0)
; SI-NEXT:    v_writelane_b32 v33, s4, 0
; SI-NEXT:    s_lshr_b32 s4, s25, 16
; SI-NEXT:    v_writelane_b32 v32, s53, 13
; SI-NEXT:    v_writelane_b32 v33, s4, 1
; SI-NEXT:    s_lshr_b32 s4, s24, 16
; SI-NEXT:    v_writelane_b32 v32, s54, 14
; SI-NEXT:    v_writelane_b32 v33, s4, 2
; SI-NEXT:    v_writelane_b32 v32, s55, 15
; SI-NEXT:    v_writelane_b32 v33, s23, 3
; SI-NEXT:    s_lshr_b32 s4, s23, 16
; SI-NEXT:    v_writelane_b32 v32, s64, 16
; SI-NEXT:    v_writelane_b32 v33, s4, 4
; SI-NEXT:    s_lshr_b32 s4, s21, 16
; SI-NEXT:    v_writelane_b32 v32, s65, 17
; SI-NEXT:    v_writelane_b32 v33, s4, 5
; SI-NEXT:    v_writelane_b32 v32, s66, 18
; SI-NEXT:    v_writelane_b32 v33, s19, 6
; SI-NEXT:    s_lshr_b32 s4, s19, 16
; SI-NEXT:    v_writelane_b32 v32, s67, 19
; SI-NEXT:    v_writelane_b32 v33, s4, 7
; SI-NEXT:    v_writelane_b32 v32, s68, 20
; SI-NEXT:    v_writelane_b32 v33, s17, 8
; SI-NEXT:    s_lshr_b32 s4, s17, 16
; SI-NEXT:    v_writelane_b32 v32, s69, 21
; SI-NEXT:    v_writelane_b32 v33, s4, 9
; SI-NEXT:    s_lshr_b32 s4, s16, 16
; SI-NEXT:    v_writelane_b32 v32, s70, 22
; SI-NEXT:    s_mov_b32 s46, s16
; SI-NEXT:    v_writelane_b32 v33, s4, 10
; SI-NEXT:    v_writelane_b32 v32, s71, 23
; SI-NEXT:    s_mov_b32 s47, s18
; SI-NEXT:    v_writelane_b32 v33, s46, 11
; SI-NEXT:    v_writelane_b32 v32, s80, 24
; SI-NEXT:    s_mov_b32 s57, s20
; SI-NEXT:    v_writelane_b32 v33, s47, 12
; SI-NEXT:    v_writelane_b32 v32, s81, 25
; SI-NEXT:    s_mov_b32 s31, s22
; SI-NEXT:    v_writelane_b32 v33, s57, 13
; SI-NEXT:    v_writelane_b32 v32, s82, 26
; SI-NEXT:    s_mov_b32 s35, s24
; SI-NEXT:    v_writelane_b32 v33, s31, 14
; SI-NEXT:    v_writelane_b32 v32, s83, 27
; SI-NEXT:    s_mov_b32 s69, s26
; SI-NEXT:    v_writelane_b32 v33, s35, 15
; SI-NEXT:    v_writelane_b32 v32, s84, 28
; SI-NEXT:    s_mov_b32 s78, s28
; SI-NEXT:    v_writelane_b32 v33, s69, 16
; SI-NEXT:    v_writelane_b32 v32, s85, 29
; SI-NEXT:    v_readfirstlane_b32 s77, v10
; SI-NEXT:    v_readfirstlane_b32 s45, v0
; SI-NEXT:    v_writelane_b32 v33, s78, 17
; SI-NEXT:    v_writelane_b32 v32, s86, 30
; SI-NEXT:    s_lshr_b32 s34, s18, 16
; SI-NEXT:    s_lshr_b32 s18, s77, 16
; SI-NEXT:    v_writelane_b32 v33, s45, 18
; SI-NEXT:    v_writelane_b32 v32, s87, 31
; SI-NEXT:    v_readfirstlane_b32 s19, v13
; SI-NEXT:    v_readfirstlane_b32 s23, v12
; SI-NEXT:    v_writelane_b32 v33, s18, 19
; SI-NEXT:    v_writelane_b32 v32, s96, 32
; SI-NEXT:    s_lshr_b32 s36, s20, 16
; SI-NEXT:    s_lshr_b32 s20, s23, 16
; SI-NEXT:    v_writelane_b32 v33, s19, 20
; SI-NEXT:    v_writelane_b32 v32, s97, 33
; SI-NEXT:    v_readfirstlane_b32 s58, v15
; SI-NEXT:    v_readfirstlane_b32 s88, v14
; SI-NEXT:    v_writelane_b32 v33, s20, 21
; SI-NEXT:    v_writelane_b32 v32, s98, 34
; SI-NEXT:    s_mov_b32 s39, s27
; SI-NEXT:    v_readfirstlane_b32 s60, v17
; SI-NEXT:    v_readfirstlane_b32 s92, v16
; SI-NEXT:    s_lshr_b32 s59, s88, 16
; SI-NEXT:    v_readfirstlane_b32 s98, v11
; SI-NEXT:    v_readfirstlane_b32 s27, v9
; SI-NEXT:    v_readfirstlane_b32 s76, v8
; SI-NEXT:    v_readfirstlane_b32 s17, v7
; SI-NEXT:    v_readfirstlane_b32 s82, v6
; SI-NEXT:    v_readfirstlane_b32 s83, v5
; SI-NEXT:    v_readfirstlane_b32 s81, v4
; SI-NEXT:    v_readfirstlane_b32 s70, v3
; SI-NEXT:    v_readfirstlane_b32 s71, v2
; SI-NEXT:    v_readfirstlane_b32 s68, v1
; SI-NEXT:    v_writelane_b32 v33, s58, 22
; SI-NEXT:    v_writelane_b32 v32, s99, 35
; SI-NEXT:    s_lshr_b32 s93, s29, 16
; SI-NEXT:    s_lshr_b32 s85, s28, 16
; SI-NEXT:    s_lshr_b32 s84, s26, 16
; SI-NEXT:    s_lshr_b32 s91, s22, 16
; SI-NEXT:    s_lshr_b32 s99, s60, 16
; SI-NEXT:    s_lshr_b32 s61, s92, 16
; SI-NEXT:    s_lshr_b32 s67, s58, 16
; SI-NEXT:    s_lshr_b32 s30, s19, 16
; SI-NEXT:    s_lshr_b32 s65, s98, 16
; SI-NEXT:    s_lshr_b32 s55, s27, 16
; SI-NEXT:    s_lshr_b32 s90, s76, 16
; SI-NEXT:    s_lshr_b32 s95, s17, 16
; SI-NEXT:    s_lshr_b32 s96, s82, 16
; SI-NEXT:    s_lshr_b32 s53, s83, 16
; SI-NEXT:    s_lshr_b32 s37, s81, 16
; SI-NEXT:    s_lshr_b32 s51, s70, 16
; SI-NEXT:    s_lshr_b32 s48, s71, 16
; SI-NEXT:    s_lshr_b32 s49, s68, 16
; SI-NEXT:    s_lshr_b32 s86, s45, 16
; SI-NEXT:    v_readfirstlane_b32 s4, v18
; SI-NEXT:    v_writelane_b32 v33, s59, 23
; SI-NEXT:    s_mov_b32 s38, s29
; SI-NEXT:    s_cmp_lg_u32 s4, 0
; SI-NEXT:    v_writelane_b32 v33, s60, 24
; SI-NEXT:    v_writelane_b32 v33, s61, 25
; SI-NEXT:    s_cbranch_scc0 .LBB111_4
; SI-NEXT:  ; %bb.1: ; %cmp.false
; SI-NEXT:    v_readlane_b32 s5, v33, 8
; SI-NEXT:    v_readlane_b32 s7, v33, 9
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s43, s5, s7
; SI-NEXT:    v_readlane_b32 s5, v33, 6
; SI-NEXT:    v_readlane_b32 s7, v33, 7
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s41, s5, s7
; SI-NEXT:    v_readlane_b32 s7, v33, 5
; SI-NEXT:    s_and_b32 s5, s21, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s15, s5, s7
; SI-NEXT:    v_readlane_b32 s5, v33, 3
; SI-NEXT:    v_readlane_b32 s7, v33, 4
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s13, s5, s7
; SI-NEXT:    v_readlane_b32 s7, v33, 1
; SI-NEXT:    s_and_b32 s5, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s11, s5, s7
; SI-NEXT:    v_readlane_b32 s7, v33, 0
; SI-NEXT:    s_and_b32 s5, s39, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    v_readlane_b32 s4, v33, 10
; SI-NEXT:    s_or_b32 s9, s5, s7
; SI-NEXT:    s_and_b32 s5, s38, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s93, 16
; SI-NEXT:    s_lshl_b32 s42, s4, 16
; SI-NEXT:    v_readlane_b32 s4, v33, 2
; SI-NEXT:    s_or_b32 s7, s5, s7
; SI-NEXT:    s_and_b32 s5, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s49, 16
; SI-NEXT:    v_writelane_b32 v33, s82, 34
; SI-NEXT:    s_or_b32 s5, s5, s44
; SI-NEXT:    s_and_b32 s56, s70, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s51, 16
; SI-NEXT:    v_writelane_b32 v33, s68, 35
; SI-NEXT:    s_or_b32 vcc_hi, s56, s44
; SI-NEXT:    s_and_b32 s56, s83, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s53, 16
; SI-NEXT:    v_writelane_b32 v33, s38, 36
; SI-NEXT:    s_mov_b32 s82, s39
; SI-NEXT:    s_or_b32 s39, s56, s44
; SI-NEXT:    s_and_b32 s56, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s95, 16
; SI-NEXT:    v_writelane_b32 v33, s76, 37
; SI-NEXT:    s_lshl_b32 s8, s84, 16
; SI-NEXT:    v_writelane_b32 v33, s84, 38
; SI-NEXT:    s_mov_b32 s84, s49
; SI-NEXT:    s_or_b32 s49, s56, s44
; SI-NEXT:    s_and_b32 s56, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s55, 16
; SI-NEXT:    s_lshl_b32 s6, s85, 16
; SI-NEXT:    v_writelane_b32 v33, s92, 39
; SI-NEXT:    s_mov_b32 s92, s77
; SI-NEXT:    s_mov_b32 s77, s70
; SI-NEXT:    s_mov_b32 s70, s85
; SI-NEXT:    s_mov_b32 s85, s51
; SI-NEXT:    s_or_b32 s51, s56, s44
; SI-NEXT:    s_and_b32 s56, s98, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s65, 16
; SI-NEXT:    s_mov_b32 s87, s21
; SI-NEXT:    s_mov_b32 s94, s53
; SI-NEXT:    s_or_b32 s53, s56, s44
; SI-NEXT:    s_and_b32 s21, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s30, 16
; SI-NEXT:    s_mov_b32 s97, s55
; SI-NEXT:    s_or_b32 s55, s21, s44
; SI-NEXT:    s_and_b32 s21, s58, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s67, 16
; SI-NEXT:    s_lshl_b32 s54, s20, 16
; SI-NEXT:    s_mov_b32 s20, s98
; SI-NEXT:    s_mov_b32 s98, s30
; SI-NEXT:    s_mov_b32 s30, s65
; SI-NEXT:    s_or_b32 s65, s21, s44
; SI-NEXT:    s_and_b32 s21, s60, 0xffff
; SI-NEXT:    s_lshl_b32 s44, s99, 16
; SI-NEXT:    s_lshl_b32 s40, s34, 16
; SI-NEXT:    s_lshl_b32 s14, s36, 16
; SI-NEXT:    s_lshl_b32 s12, s91, 16
; SI-NEXT:    s_lshl_b32 s10, s4, 16
; SI-NEXT:    s_lshl_b32 s4, s86, 16
; SI-NEXT:    s_mov_b32 s80, s86
; SI-NEXT:    s_mov_b32 s86, s91
; SI-NEXT:    s_mov_b32 s91, s36
; SI-NEXT:    s_mov_b32 s36, s34
; SI-NEXT:    s_mov_b32 s34, s99
; SI-NEXT:    s_mov_b32 s99, s67
; SI-NEXT:    s_or_b32 s67, s21, s44
; SI-NEXT:    s_and_b32 s21, s46, 0xffff
; SI-NEXT:    s_or_b32 s62, s21, s42
; SI-NEXT:    s_and_b32 s21, s47, 0xffff
; SI-NEXT:    s_or_b32 s74, s21, s40
; SI-NEXT:    s_and_b32 s21, s57, 0xffff
; SI-NEXT:    v_writelane_b32 v33, s23, 40
; SI-NEXT:    s_or_b32 s72, s21, s14
; SI-NEXT:    s_lshr_b64 s[22:23], s[14:15], 16
; SI-NEXT:    s_and_b32 s14, s31, 0xffff
; SI-NEXT:    s_mov_b32 s68, s25
; SI-NEXT:    s_or_b32 s60, s14, s12
; SI-NEXT:    s_lshr_b64 s[24:25], s[12:13], 16
; SI-NEXT:    s_and_b32 s12, s35, 0xffff
; SI-NEXT:    s_mov_b32 s89, s27
; SI-NEXT:    s_or_b32 s58, s12, s10
; SI-NEXT:    s_lshr_b64 s[26:27], s[10:11], 16
; SI-NEXT:    s_and_b32 s10, s69, 0xffff
; SI-NEXT:    s_or_b32 s56, s10, s8
; SI-NEXT:    s_lshr_b64 s[28:29], s[8:9], 16
; SI-NEXT:    s_and_b32 s8, s78, 0xffff
; SI-NEXT:    s_or_b32 s46, s8, s6
; SI-NEXT:    s_lshr_b64 s[78:79], s[6:7], 16
; SI-NEXT:    s_and_b32 s6, s45, 0xffff
; SI-NEXT:    v_writelane_b32 v33, s17, 41
; SI-NEXT:    s_or_b32 s44, s6, s4
; SI-NEXT:    s_mov_b32 s45, s5
; SI-NEXT:    s_lshr_b64 s[4:5], s[4:5], 16
; SI-NEXT:    s_lshl_b32 vcc_lo, s48, 16
; SI-NEXT:    v_writelane_b32 v33, s4, 26
; SI-NEXT:    s_mov_b32 s47, s7
; SI-NEXT:    v_writelane_b32 v33, s5, 27
; SI-NEXT:    s_and_b32 s4, s71, 0xffff
; SI-NEXT:    s_lshr_b64 s[6:7], vcc, 16
; SI-NEXT:    s_lshl_b32 s38, s37, 16
; SI-NEXT:    s_lshl_b32 s52, s18, 16
; SI-NEXT:    s_lshr_b64 s[18:19], s[42:43], 16
; SI-NEXT:    s_or_b32 s42, s4, vcc_lo
; SI-NEXT:    v_writelane_b32 v33, s6, 28
; SI-NEXT:    s_and_b32 s4, s81, 0xffff
; SI-NEXT:    s_lshr_b64 s[16:17], s[40:41], 16
; SI-NEXT:    v_writelane_b32 v33, s7, 29
; SI-NEXT:    s_or_b32 s40, s4, s38
; SI-NEXT:    s_lshr_b64 s[4:5], s[38:39], 16
; SI-NEXT:    v_writelane_b32 v33, s4, 30
; SI-NEXT:    v_writelane_b32 v33, s5, 31
; SI-NEXT:    s_mov_b32 s75, s41
; SI-NEXT:    s_mov_b32 s41, s39
; SI-NEXT:    s_mov_b32 s39, s82
; SI-NEXT:    v_readlane_b32 s82, v33, 34
; SI-NEXT:    s_mov_b32 s76, s48
; SI-NEXT:    s_lshl_b32 s48, s96, 16
; SI-NEXT:    s_and_b32 s4, s82, 0xffff
; SI-NEXT:    s_or_b32 s14, s4, s48
; SI-NEXT:    s_lshr_b64 s[4:5], s[48:49], 16
; SI-NEXT:    s_mov_b32 s25, s68
; SI-NEXT:    v_readlane_b32 s38, v33, 36
; SI-NEXT:    v_readlane_b32 s68, v33, 35
; SI-NEXT:    v_writelane_b32 v33, s4, 32
; SI-NEXT:    v_writelane_b32 v33, s5, 33
; SI-NEXT:    s_mov_b32 s48, s76
; SI-NEXT:    v_readlane_b32 s76, v33, 37
; SI-NEXT:    s_lshl_b32 s50, s90, 16
; SI-NEXT:    s_and_b32 s4, s76, 0xffff
; SI-NEXT:    s_lshl_b32 s66, s61, 16
; SI-NEXT:    s_mov_b32 s61, s13
; SI-NEXT:    s_or_b32 s12, s4, s50
; SI-NEXT:    s_mov_b32 s13, s51
; SI-NEXT:    s_lshr_b64 s[50:51], s[50:51], 16
; SI-NEXT:    s_mov_b32 s51, s85
; SI-NEXT:    s_mov_b32 s85, s70
; SI-NEXT:    s_mov_b32 s70, s77
; SI-NEXT:    s_mov_b32 s77, s92
; SI-NEXT:    v_readlane_b32 s23, v33, 40
; SI-NEXT:    s_and_b32 s4, s77, 0xffff
; SI-NEXT:    s_or_b32 s10, s4, s52
; SI-NEXT:    s_and_b32 s4, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s64, s59, 16
; SI-NEXT:    v_readlane_b32 s92, v33, 39
; SI-NEXT:    s_or_b32 s8, s4, s54
; SI-NEXT:    s_and_b32 s4, s88, 0xffff
; SI-NEXT:    s_or_b32 s6, s4, s64
; SI-NEXT:    s_and_b32 s4, s92, 0xffff
; SI-NEXT:    s_mov_b32 s59, s11
; SI-NEXT:    s_mov_b32 s57, s9
; SI-NEXT:    s_mov_b32 s11, s53
; SI-NEXT:    s_lshr_b64 s[52:53], s[52:53], 16
; SI-NEXT:    s_mov_b32 s9, s55
; SI-NEXT:    s_lshr_b64 s[54:55], s[54:55], 16
; SI-NEXT:    s_mov_b32 s7, s65
; SI-NEXT:    s_lshr_b64 s[64:65], s[64:65], 16
; SI-NEXT:    s_or_b32 s4, s4, s66
; SI-NEXT:    s_mov_b32 s5, s67
; SI-NEXT:    s_lshr_b64 s[66:67], s[66:67], 16
; SI-NEXT:    s_mov_b32 s63, s43
; SI-NEXT:    s_mov_b32 s73, s15
; SI-NEXT:    s_mov_b32 s43, vcc_hi
; SI-NEXT:    s_mov_b32 s15, s49
; SI-NEXT:    s_mov_b32 s49, s84
; SI-NEXT:    v_readlane_b32 s84, v33, 38
; SI-NEXT:    s_mov_b32 s53, s94
; SI-NEXT:    s_mov_b32 s21, s87
; SI-NEXT:    v_readlane_b32 s17, v33, 41
; SI-NEXT:    s_mov_b32 s55, s97
; SI-NEXT:    s_mov_b32 s65, s30
; SI-NEXT:    s_mov_b32 s30, s98
; SI-NEXT:    s_mov_b32 s98, s20
; SI-NEXT:    s_mov_b32 s67, s99
; SI-NEXT:    s_mov_b32 s99, s34
; SI-NEXT:    s_mov_b32 s34, s36
; SI-NEXT:    s_mov_b32 s36, s91
; SI-NEXT:    s_mov_b32 s91, s86
; SI-NEXT:    s_mov_b32 s86, s80
; SI-NEXT:    s_mov_b32 s27, s89
; SI-NEXT:    s_mov_b32 s20, s16
; SI-NEXT:    s_cbranch_execnz .LBB111_3
; SI-NEXT:  .LBB111_2: ; %cmp.true
; SI-NEXT:    s_add_i32 s20, s92, 3
; SI-NEXT:    v_readlane_b32 s5, v33, 25
; SI-NEXT:    s_and_b32 s4, s20, 0xffff
; SI-NEXT:    s_lshl_b32 s5, s5, 16
; SI-NEXT:    s_or_b32 s4, s5, s4
; SI-NEXT:    v_readlane_b32 s5, v33, 24
; SI-NEXT:    s_add_i32 s89, s5, 3
; SI-NEXT:    s_and_b32 s5, s89, 0xffff
; SI-NEXT:    s_lshl_b32 s6, s99, 16
; SI-NEXT:    s_add_i32 s88, s88, 3
; SI-NEXT:    v_readlane_b32 s7, v33, 23
; SI-NEXT:    s_or_b32 s5, s6, s5
; SI-NEXT:    s_and_b32 s6, s88, 0xffff
; SI-NEXT:    s_lshl_b32 s7, s7, 16
; SI-NEXT:    s_or_b32 s6, s7, s6
; SI-NEXT:    v_readlane_b32 s7, v33, 22
; SI-NEXT:    s_add_i32 s19, s7, 3
; SI-NEXT:    s_and_b32 s7, s19, 0xffff
; SI-NEXT:    s_lshl_b32 s8, s67, 16
; SI-NEXT:    s_add_i32 s18, s23, 3
; SI-NEXT:    v_readlane_b32 s9, v33, 21
; SI-NEXT:    s_or_b32 s7, s8, s7
; SI-NEXT:    s_and_b32 s8, s18, 0xffff
; SI-NEXT:    s_lshl_b32 s9, s9, 16
; SI-NEXT:    s_or_b32 s8, s9, s8
; SI-NEXT:    v_readlane_b32 s9, v33, 20
; SI-NEXT:    s_add_i32 s79, s9, 3
; SI-NEXT:    s_and_b32 s9, s79, 0xffff
; SI-NEXT:    s_lshl_b32 s10, s30, 16
; SI-NEXT:    s_add_i32 s78, s77, 3
; SI-NEXT:    v_readlane_b32 s11, v33, 19
; SI-NEXT:    s_or_b32 s9, s10, s9
; SI-NEXT:    s_and_b32 s10, s78, 0xffff
; SI-NEXT:    s_lshl_b32 s11, s11, 16
; SI-NEXT:    s_add_i32 s77, s98, 3
; SI-NEXT:    s_or_b32 s10, s11, s10
; SI-NEXT:    s_and_b32 s11, s77, 0xffff
; SI-NEXT:    s_lshl_b32 s12, s65, 16
; SI-NEXT:    s_add_i32 s76, s76, 3
; SI-NEXT:    s_or_b32 s11, s12, s11
; SI-NEXT:    s_and_b32 s12, s76, 0xffff
; SI-NEXT:    s_lshl_b32 s13, s90, 16
; SI-NEXT:    s_add_i32 s16, s27, 3
; SI-NEXT:    s_or_b32 s12, s13, s12
; SI-NEXT:    s_and_b32 s13, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s14, s55, 16
; SI-NEXT:    s_add_i32 s82, s82, 3
; SI-NEXT:    s_or_b32 s13, s14, s13
; SI-NEXT:    s_and_b32 s14, s82, 0xffff
; SI-NEXT:    s_lshl_b32 s15, s96, 16
; SI-NEXT:    s_add_i32 s17, s17, 3
; SI-NEXT:    s_or_b32 s14, s15, s14
; SI-NEXT:    s_and_b32 s15, s17, 0xffff
; SI-NEXT:    s_lshl_b32 s16, s95, 16
; SI-NEXT:    s_add_i32 s81, s81, 3
; SI-NEXT:    s_or_b32 s15, s16, s15
; SI-NEXT:    s_and_b32 s16, s81, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s37, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s80, s83, 3
; SI-NEXT:    s_add_i32 s40, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s80, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s53, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s71, s71, 3
; SI-NEXT:    s_add_i32 s41, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s71, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s48, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s70, s70, 3
; SI-NEXT:    s_add_i32 s42, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s70, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s51, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s43, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 18
; SI-NEXT:    s_add_i32 s69, s16, 3
; SI-NEXT:    s_and_b32 s16, s69, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s86, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s68, s68, 3
; SI-NEXT:    s_add_i32 s44, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s68, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s49, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s45, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 17
; SI-NEXT:    s_add_i32 s28, s16, 3
; SI-NEXT:    s_and_b32 s16, s28, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s85, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s29, s38, 3
; SI-NEXT:    s_add_i32 s46, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s29, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s93, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s47, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 16
; SI-NEXT:    s_add_i32 s26, s16, 3
; SI-NEXT:    s_and_b32 s16, s26, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s84, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s27, s39, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 0
; SI-NEXT:    s_add_i32 s56, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s27, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s57, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 15
; SI-NEXT:    s_add_i32 s24, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 2
; SI-NEXT:    s_and_b32 s16, s24, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s25, s25, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 1
; SI-NEXT:    s_add_i32 s58, s16, 0x30000
; SI-NEXT:    s_and_b32 s16, s25, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s59, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 14
; SI-NEXT:    s_add_i32 s22, s16, 3
; SI-NEXT:    s_and_b32 s16, s22, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s91, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s60, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 3
; SI-NEXT:    s_add_i32 s23, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 4
; SI-NEXT:    s_and_b32 s16, s23, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s61, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 13
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s36, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s72, s16, 0x30000
; SI-NEXT:    s_add_i32 s16, s21, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 5
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s73, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 12
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s34, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s74, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 6
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 7
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s75, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 11
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 10
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s62, s16, 0x30000
; SI-NEXT:    v_readlane_b32 s16, v33, 8
; SI-NEXT:    s_add_i32 s16, s16, 3
; SI-NEXT:    v_readlane_b32 s17, v33, 9
; SI-NEXT:    s_and_b32 s16, s16, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s17, 16
; SI-NEXT:    s_or_b32 s16, s17, s16
; SI-NEXT:    s_add_i32 s63, s16, 0x30000
; SI-NEXT:    s_lshr_b64 s[16:17], s[44:45], 16
; SI-NEXT:    v_writelane_b32 v33, s16, 26
; SI-NEXT:    v_writelane_b32 v33, s17, 27
; SI-NEXT:    s_lshr_b64 s[16:17], s[42:43], 16
; SI-NEXT:    v_writelane_b32 v33, s16, 28
; SI-NEXT:    v_writelane_b32 v33, s17, 29
; SI-NEXT:    s_lshr_b64 s[16:17], s[40:41], 16
; SI-NEXT:    s_add_i32 s14, s14, 0x30000
; SI-NEXT:    s_add_i32 s15, s15, 0x30000
; SI-NEXT:    v_writelane_b32 v33, s16, 30
; SI-NEXT:    v_writelane_b32 v33, s17, 31
; SI-NEXT:    s_lshr_b64 s[16:17], s[14:15], 16
; SI-NEXT:    v_writelane_b32 v33, s16, 32
; SI-NEXT:    v_writelane_b32 v33, s17, 33
; SI-NEXT:    s_lshr_b32 s16, s63, 16
; SI-NEXT:    v_writelane_b32 v33, s16, 9
; SI-NEXT:    s_lshr_b32 s16, s75, 16
; SI-NEXT:    v_writelane_b32 v33, s16, 7
; SI-NEXT:    s_lshr_b32 s16, s73, 16
; SI-NEXT:    s_add_i32 s4, s4, 0x30000
; SI-NEXT:    s_add_i32 s5, s5, 0x30000
; SI-NEXT:    s_add_i32 s6, s6, 0x30000
; SI-NEXT:    s_add_i32 s7, s7, 0x30000
; SI-NEXT:    s_add_i32 s8, s8, 0x30000
; SI-NEXT:    s_add_i32 s9, s9, 0x30000
; SI-NEXT:    s_add_i32 s10, s10, 0x30000
; SI-NEXT:    s_add_i32 s11, s11, 0x30000
; SI-NEXT:    s_add_i32 s12, s12, 0x30000
; SI-NEXT:    s_add_i32 s13, s13, 0x30000
; SI-NEXT:    v_writelane_b32 v33, s16, 5
; SI-NEXT:    s_lshr_b32 s16, s61, 16
; SI-NEXT:    s_lshr_b64 s[50:51], s[12:13], 16
; SI-NEXT:    s_lshr_b64 s[52:53], s[10:11], 16
; SI-NEXT:    s_lshr_b64 s[54:55], s[8:9], 16
; SI-NEXT:    s_lshr_b64 s[64:65], s[6:7], 16
; SI-NEXT:    s_lshr_b64 s[66:67], s[4:5], 16
; SI-NEXT:    v_writelane_b32 v33, s16, 4
; SI-NEXT:    s_lshr_b32 s16, s59, 16
; SI-NEXT:    s_lshr_b64 s[18:19], s[62:63], 16
; SI-NEXT:    s_lshr_b64 s[20:21], s[74:75], 16
; SI-NEXT:    s_lshr_b64 s[22:23], s[72:73], 16
; SI-NEXT:    s_lshr_b64 s[24:25], s[60:61], 16
; SI-NEXT:    s_lshr_b64 s[26:27], s[58:59], 16
; SI-NEXT:    s_lshr_b64 s[28:29], s[56:57], 16
; SI-NEXT:    s_lshr_b64 s[78:79], s[46:47], 16
; SI-NEXT:    v_writelane_b32 v33, s16, 1
; SI-NEXT:    s_lshr_b32 s16, s57, 16
; SI-NEXT:    s_lshr_b32 s93, s47, 16
; SI-NEXT:    s_lshr_b32 s49, s45, 16
; SI-NEXT:    s_lshr_b32 s51, s43, 16
; SI-NEXT:    s_lshr_b32 s53, s41, 16
; SI-NEXT:    s_lshr_b32 s95, s15, 16
; SI-NEXT:    s_lshr_b32 s55, s13, 16
; SI-NEXT:    s_lshr_b32 s65, s11, 16
; SI-NEXT:    s_lshr_b32 s30, s9, 16
; SI-NEXT:    s_lshr_b32 s67, s7, 16
; SI-NEXT:    s_lshr_b32 s99, s5, 16
; SI-NEXT:    v_writelane_b32 v33, s16, 0
; SI-NEXT:  .LBB111_3: ; %end
; SI-NEXT:    s_and_b32 s16, s62, 0xffff
; SI-NEXT:    s_lshl_b32 s17, s18, 16
; SI-NEXT:    v_readlane_b32 s18, v33, 9
; SI-NEXT:    s_or_b32 s16, s16, s17
; SI-NEXT:    s_and_b32 s17, s63, 0xffff
; SI-NEXT:    s_lshl_b32 s18, s18, 16
; SI-NEXT:    s_or_b32 s17, s17, s18
; SI-NEXT:    s_and_b32 s18, s74, 0xffff
; SI-NEXT:    s_lshl_b32 s19, s20, 16
; SI-NEXT:    v_readlane_b32 s20, v33, 7
; SI-NEXT:    s_or_b32 s18, s18, s19
; SI-NEXT:    s_and_b32 s19, s75, 0xffff
; SI-NEXT:    s_lshl_b32 s20, s20, 16
; SI-NEXT:    s_or_b32 s19, s19, s20
; SI-NEXT:    s_and_b32 s20, s72, 0xffff
; SI-NEXT:    s_lshl_b32 s21, s22, 16
; SI-NEXT:    v_readlane_b32 s22, v33, 5
; SI-NEXT:    s_or_b32 s20, s20, s21
; SI-NEXT:    s_and_b32 s21, s73, 0xffff
; SI-NEXT:    s_lshl_b32 s22, s22, 16
; SI-NEXT:    s_or_b32 s21, s21, s22
; SI-NEXT:    s_and_b32 s22, s60, 0xffff
; SI-NEXT:    s_lshl_b32 s23, s24, 16
; SI-NEXT:    v_readlane_b32 s24, v33, 4
; SI-NEXT:    s_or_b32 s22, s22, s23
; SI-NEXT:    s_and_b32 s23, s61, 0xffff
; SI-NEXT:    s_lshl_b32 s24, s24, 16
; SI-NEXT:    s_or_b32 s23, s23, s24
; SI-NEXT:    s_and_b32 s24, s58, 0xffff
; SI-NEXT:    s_lshl_b32 s25, s26, 16
; SI-NEXT:    v_readlane_b32 s26, v33, 1
; SI-NEXT:    s_or_b32 s24, s24, s25
; SI-NEXT:    s_and_b32 s25, s59, 0xffff
; SI-NEXT:    s_lshl_b32 s26, s26, 16
; SI-NEXT:    s_or_b32 s25, s25, s26
; SI-NEXT:    s_and_b32 s26, s56, 0xffff
; SI-NEXT:    s_lshl_b32 s27, s28, 16
; SI-NEXT:    v_readlane_b32 s28, v33, 0
; SI-NEXT:    s_or_b32 s26, s26, s27
; SI-NEXT:    s_and_b32 s27, s57, 0xffff
; SI-NEXT:    s_lshl_b32 s28, s28, 16
; SI-NEXT:    s_or_b32 s27, s27, s28
; SI-NEXT:    s_and_b32 s28, s46, 0xffff
; SI-NEXT:    s_lshl_b32 s29, s78, 16
; SI-NEXT:    s_or_b32 s28, s28, s29
; SI-NEXT:    s_and_b32 s29, s47, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s93, 16
; SI-NEXT:    s_or_b32 s29, s29, s46
; SI-NEXT:    v_readlane_b32 s46, v33, 26
; SI-NEXT:    s_and_b32 s44, s44, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    v_readlane_b32 s47, v33, 27
; SI-NEXT:    s_or_b32 s44, s44, s46
; SI-NEXT:    s_and_b32 s45, s45, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s49, 16
; SI-NEXT:    s_or_b32 s45, s45, s46
; SI-NEXT:    v_readlane_b32 s46, v33, 28
; SI-NEXT:    s_and_b32 s42, s42, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    v_readlane_b32 s47, v33, 29
; SI-NEXT:    s_or_b32 s42, s42, s46
; SI-NEXT:    s_and_b32 s43, s43, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s51, 16
; SI-NEXT:    s_or_b32 s43, s43, s46
; SI-NEXT:    v_readlane_b32 s46, v33, 30
; SI-NEXT:    s_and_b32 s40, s40, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    v_readlane_b32 s47, v33, 31
; SI-NEXT:    s_or_b32 s40, s40, s46
; SI-NEXT:    s_and_b32 s41, s41, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s53, 16
; SI-NEXT:    s_or_b32 s41, s41, s46
; SI-NEXT:    v_readlane_b32 s46, v33, 32
; SI-NEXT:    s_and_b32 s14, s14, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s46, 16
; SI-NEXT:    s_or_b32 s14, s14, s46
; SI-NEXT:    s_and_b32 s15, s15, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s95, 16
; SI-NEXT:    s_or_b32 s15, s15, s46
; SI-NEXT:    s_and_b32 s12, s12, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s50, 16
; SI-NEXT:    s_or_b32 s12, s12, s46
; SI-NEXT:    s_and_b32 s13, s13, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s55, 16
; SI-NEXT:    s_or_b32 s13, s13, s46
; SI-NEXT:    s_and_b32 s10, s10, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s52, 16
; SI-NEXT:    s_or_b32 s10, s10, s46
; SI-NEXT:    s_and_b32 s11, s11, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s65, 16
; SI-NEXT:    s_or_b32 s11, s11, s46
; SI-NEXT:    s_and_b32 s8, s8, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s54, 16
; SI-NEXT:    s_or_b32 s8, s8, s46
; SI-NEXT:    s_and_b32 s9, s9, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s30, 16
; SI-NEXT:    s_or_b32 s9, s9, s46
; SI-NEXT:    s_and_b32 s6, s6, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s64, 16
; SI-NEXT:    s_or_b32 s6, s6, s46
; SI-NEXT:    s_and_b32 s7, s7, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s67, 16
; SI-NEXT:    s_or_b32 s7, s7, s46
; SI-NEXT:    s_and_b32 s4, s4, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s66, 16
; SI-NEXT:    s_or_b32 s4, s4, s46
; SI-NEXT:    s_and_b32 s5, s5, 0xffff
; SI-NEXT:    s_lshl_b32 s46, s99, 16
; SI-NEXT:    s_or_b32 s5, s5, s46
; SI-NEXT:    v_readlane_b32 s47, v33, 33
; SI-NEXT:    v_mov_b32_e32 v0, s16
; SI-NEXT:    v_mov_b32_e32 v1, s17
; SI-NEXT:    v_mov_b32_e32 v2, s18
; SI-NEXT:    v_mov_b32_e32 v3, s19
; SI-NEXT:    v_mov_b32_e32 v4, s20
; SI-NEXT:    v_mov_b32_e32 v5, s21
; SI-NEXT:    v_mov_b32_e32 v6, s22
; SI-NEXT:    v_mov_b32_e32 v7, s23
; SI-NEXT:    v_mov_b32_e32 v8, s24
; SI-NEXT:    v_mov_b32_e32 v9, s25
; SI-NEXT:    v_mov_b32_e32 v10, s26
; SI-NEXT:    v_mov_b32_e32 v11, s27
; SI-NEXT:    v_mov_b32_e32 v12, s28
; SI-NEXT:    v_mov_b32_e32 v13, s29
; SI-NEXT:    v_mov_b32_e32 v14, s44
; SI-NEXT:    v_mov_b32_e32 v15, s45
; SI-NEXT:    v_mov_b32_e32 v16, s42
; SI-NEXT:    v_mov_b32_e32 v17, s43
; SI-NEXT:    v_mov_b32_e32 v18, s40
; SI-NEXT:    v_mov_b32_e32 v19, s41
; SI-NEXT:    v_mov_b32_e32 v20, s14
; SI-NEXT:    v_mov_b32_e32 v21, s15
; SI-NEXT:    v_mov_b32_e32 v22, s12
; SI-NEXT:    v_mov_b32_e32 v23, s13
; SI-NEXT:    v_mov_b32_e32 v24, s10
; SI-NEXT:    v_mov_b32_e32 v25, s11
; SI-NEXT:    v_mov_b32_e32 v26, s8
; SI-NEXT:    v_mov_b32_e32 v27, s9
; SI-NEXT:    v_mov_b32_e32 v28, s6
; SI-NEXT:    v_mov_b32_e32 v29, s7
; SI-NEXT:    v_mov_b32_e32 v30, s4
; SI-NEXT:    v_mov_b32_e32 v31, s5
; SI-NEXT:    v_readlane_b32 s99, v32, 35
; SI-NEXT:    v_readlane_b32 s98, v32, 34
; SI-NEXT:    v_readlane_b32 s97, v32, 33
; SI-NEXT:    v_readlane_b32 s96, v32, 32
; SI-NEXT:    v_readlane_b32 s87, v32, 31
; SI-NEXT:    v_readlane_b32 s86, v32, 30
; SI-NEXT:    v_readlane_b32 s85, v32, 29
; SI-NEXT:    v_readlane_b32 s84, v32, 28
; SI-NEXT:    v_readlane_b32 s83, v32, 27
; SI-NEXT:    v_readlane_b32 s82, v32, 26
; SI-NEXT:    v_readlane_b32 s81, v32, 25
; SI-NEXT:    v_readlane_b32 s80, v32, 24
; SI-NEXT:    v_readlane_b32 s71, v32, 23
; SI-NEXT:    v_readlane_b32 s70, v32, 22
; SI-NEXT:    v_readlane_b32 s69, v32, 21
; SI-NEXT:    v_readlane_b32 s68, v32, 20
; SI-NEXT:    v_readlane_b32 s67, v32, 19
; SI-NEXT:    v_readlane_b32 s66, v32, 18
; SI-NEXT:    v_readlane_b32 s65, v32, 17
; SI-NEXT:    v_readlane_b32 s64, v32, 16
; SI-NEXT:    v_readlane_b32 s55, v32, 15
; SI-NEXT:    v_readlane_b32 s54, v32, 14
; SI-NEXT:    v_readlane_b32 s53, v32, 13
; SI-NEXT:    v_readlane_b32 s52, v32, 12
; SI-NEXT:    v_readlane_b32 s51, v32, 11
; SI-NEXT:    v_readlane_b32 s50, v32, 10
; SI-NEXT:    v_readlane_b32 s49, v32, 9
; SI-NEXT:    v_readlane_b32 s48, v32, 8
; SI-NEXT:    v_readlane_b32 s39, v32, 7
; SI-NEXT:    v_readlane_b32 s38, v32, 6
; SI-NEXT:    v_readlane_b32 s37, v32, 5
; SI-NEXT:    v_readlane_b32 s36, v32, 4
; SI-NEXT:    v_readlane_b32 s35, v32, 3
; SI-NEXT:    v_readlane_b32 s34, v32, 2
; SI-NEXT:    v_readlane_b32 s31, v32, 1
; SI-NEXT:    v_readlane_b32 s30, v32, 0
; SI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; SI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; SI-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
; SI-NEXT:    s_mov_b64 exec, s[4:5]
; SI-NEXT:    s_waitcnt vmcnt(0)
; SI-NEXT:    s_setpc_b64 s[30:31]
; SI-NEXT:  .LBB111_4:
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v33, s4, 26
; SI-NEXT:    v_writelane_b32 v33, s5, 27
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    ; implicit-def: $sgpr62
; SI-NEXT:    ; implicit-def: $sgpr18
; SI-NEXT:    ; implicit-def: $sgpr74
; SI-NEXT:    ; implicit-def: $sgpr20
; SI-NEXT:    ; implicit-def: $sgpr72
; SI-NEXT:    ; implicit-def: $sgpr22
; SI-NEXT:    ; implicit-def: $sgpr60
; SI-NEXT:    ; implicit-def: $sgpr24
; SI-NEXT:    ; implicit-def: $sgpr58
; SI-NEXT:    ; implicit-def: $sgpr26
; SI-NEXT:    ; implicit-def: $sgpr56
; SI-NEXT:    ; implicit-def: $sgpr28
; SI-NEXT:    ; implicit-def: $sgpr46
; SI-NEXT:    ; implicit-def: $sgpr78
; SI-NEXT:    ; implicit-def: $sgpr44
; SI-NEXT:    ; implicit-def: $sgpr42
; SI-NEXT:    ; implicit-def: $sgpr40
; SI-NEXT:    ; implicit-def: $sgpr14
; SI-NEXT:    ; implicit-def: $sgpr12
; SI-NEXT:    ; implicit-def: $sgpr50
; SI-NEXT:    ; implicit-def: $sgpr10
; SI-NEXT:    ; implicit-def: $sgpr52
; SI-NEXT:    ; implicit-def: $sgpr8
; SI-NEXT:    ; implicit-def: $sgpr54
; SI-NEXT:    ; implicit-def: $sgpr6
; SI-NEXT:    ; implicit-def: $sgpr64
; SI-NEXT:    ; implicit-def: $sgpr66
; SI-NEXT:    v_writelane_b32 v33, s4, 28
; SI-NEXT:    v_writelane_b32 v33, s5, 29
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v33, s4, 30
; SI-NEXT:    v_writelane_b32 v33, s5, 31
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    v_writelane_b32 v33, s4, 32
; SI-NEXT:    v_writelane_b32 v33, s5, 33
; SI-NEXT:    ; implicit-def: $sgpr4
; SI-NEXT:    s_branch .LBB111_2
;
; VI-LABEL: bitcast_v64i16_to_v64f16_scalar:
; VI:       ; %bb.0:
; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    v_writelane_b32 v32, s30, 0
; VI-NEXT:    v_writelane_b32 v32, s31, 1
; VI-NEXT:    v_writelane_b32 v32, s34, 2
; VI-NEXT:    v_writelane_b32 v32, s35, 3
; VI-NEXT:    v_writelane_b32 v32, s36, 4
; VI-NEXT:    v_writelane_b32 v32, s37, 5
; VI-NEXT:    v_readfirstlane_b32 s4, v18
; VI-NEXT:    v_writelane_b32 v32, s38, 6
; VI-NEXT:    v_readfirstlane_b32 s8, v17
; VI-NEXT:    v_readfirstlane_b32 s9, v16
; VI-NEXT:    v_readfirstlane_b32 s10, v15
; VI-NEXT:    v_readfirstlane_b32 s11, v14
; VI-NEXT:    v_readfirstlane_b32 s12, v13
; VI-NEXT:    v_readfirstlane_b32 s13, v12
; VI-NEXT:    v_readfirstlane_b32 s14, v11
; VI-NEXT:    v_readfirstlane_b32 s15, v10
; VI-NEXT:    v_readfirstlane_b32 s40, v9
; VI-NEXT:    v_readfirstlane_b32 s41, v8
; VI-NEXT:    v_readfirstlane_b32 s42, v7
; VI-NEXT:    v_readfirstlane_b32 s43, v6
; VI-NEXT:    v_readfirstlane_b32 s44, v5
; VI-NEXT:    v_readfirstlane_b32 s45, v4
; VI-NEXT:    v_readfirstlane_b32 s46, v3
; VI-NEXT:    v_readfirstlane_b32 s47, v2
; VI-NEXT:    v_readfirstlane_b32 s6, v1
; VI-NEXT:    s_cmp_lg_u32 s4, 0
; VI-NEXT:    v_readfirstlane_b32 s7, v0
; VI-NEXT:    v_writelane_b32 v32, s39, 7
; VI-NEXT:    s_cbranch_scc0 .LBB111_4
; VI-NEXT:  ; %bb.1: ; %cmp.false
; VI-NEXT:    s_cbranch_execnz .LBB111_3
; VI-NEXT:  .LBB111_2: ; %cmp.true
; VI-NEXT:    s_and_b32 s4, s47, 0xffff0000
; VI-NEXT:    s_add_i32 s5, s47, 3
; VI-NEXT:    s_and_b32 s47, s46, 0xffff0000
; VI-NEXT:    s_add_i32 s46, s46, 3
; VI-NEXT:    s_and_b32 s56, s45, 0xffff0000
; VI-NEXT:    s_add_i32 s45, s45, 3
; VI-NEXT:    s_and_b32 s57, s44, 0xffff0000
; VI-NEXT:    s_add_i32 s44, s44, 3
; VI-NEXT:    s_and_b32 s58, s43, 0xffff0000
; VI-NEXT:    s_add_i32 s43, s43, 3
; VI-NEXT:    s_and_b32 s59, s42, 0xffff0000
; VI-NEXT:    s_add_i32 s42, s42, 3
; VI-NEXT:    s_and_b32 s60, s41, 0xffff0000
; VI-NEXT:    s_add_i32 s41, s41, 3
; VI-NEXT:    s_and_b32 s61, s40, 0xffff0000
; VI-NEXT:    s_add_i32 s40, s40, 3
; VI-NEXT:    s_and_b32 s62, s15, 0xffff0000
; VI-NEXT:    s_add_i32 s15, s15, 3
; VI-NEXT:    s_and_b32 s63, s14, 0xffff0000
; VI-NEXT:    s_add_i32 s14, s14, 3
; VI-NEXT:    s_and_b32 s72, s13, 0xffff0000
; VI-NEXT:    s_add_i32 s13, s13, 3
; VI-NEXT:    s_and_b32 s73, s12, 0xffff0000
; VI-NEXT:    s_add_i32 s12, s12, 3
; VI-NEXT:    s_and_b32 s74, s11, 0xffff0000
; VI-NEXT:    s_add_i32 s11, s11, 3
; VI-NEXT:    s_and_b32 s75, s10, 0xffff0000
; VI-NEXT:    s_add_i32 s10, s10, 3
; VI-NEXT:    s_and_b32 s76, s9, 0xffff0000
; VI-NEXT:    s_add_i32 s9, s9, 3
; VI-NEXT:    s_and_b32 s77, s8, 0xffff0000
; VI-NEXT:    s_add_i32 s8, s8, 3
; VI-NEXT:    s_and_b32 s78, s16, 0xffff0000
; VI-NEXT:    s_add_i32 s16, s16, 3
; VI-NEXT:    s_and_b32 s79, s17, 0xffff0000
; VI-NEXT:    s_add_i32 s17, s17, 3
; VI-NEXT:    s_and_b32 s88, s18, 0xffff0000
; VI-NEXT:    s_add_i32 s18, s18, 3
; VI-NEXT:    s_and_b32 s89, s19, 0xffff0000
; VI-NEXT:    s_add_i32 s19, s19, 3
; VI-NEXT:    s_and_b32 s90, s20, 0xffff0000
; VI-NEXT:    s_add_i32 s20, s20, 3
; VI-NEXT:    s_and_b32 s91, s21, 0xffff0000
; VI-NEXT:    s_add_i32 s21, s21, 3
; VI-NEXT:    s_and_b32 vcc_lo, s22, 0xffff0000
; VI-NEXT:    s_add_i32 s22, s22, 3
; VI-NEXT:    s_and_b32 vcc_hi, s23, 0xffff0000
; VI-NEXT:    s_add_i32 s23, s23, 3
; VI-NEXT:    s_and_b32 s30, s24, 0xffff0000
; VI-NEXT:    s_add_i32 s24, s24, 3
; VI-NEXT:    s_and_b32 s31, s25, 0xffff0000
; VI-NEXT:    s_add_i32 s25, s25, 3
; VI-NEXT:    s_and_b32 s34, s26, 0xffff0000
; VI-NEXT:    s_add_i32 s26, s26, 3
; VI-NEXT:    s_and_b32 s35, s27, 0xffff0000
; VI-NEXT:    s_add_i32 s27, s27, 3
; VI-NEXT:    s_and_b32 s36, s28, 0xffff0000
; VI-NEXT:    s_add_i32 s28, s28, 3
; VI-NEXT:    s_and_b32 s37, s29, 0xffff0000
; VI-NEXT:    s_add_i32 s29, s29, 3
; VI-NEXT:    s_and_b32 s38, s7, 0xffff0000
; VI-NEXT:    s_add_i32 s7, s7, 3
; VI-NEXT:    s_and_b32 s39, s6, 0xffff0000
; VI-NEXT:    s_add_i32 s6, s6, 3
; VI-NEXT:    s_and_b32 s6, s6, 0xffff
; VI-NEXT:    s_and_b32 s7, s7, 0xffff
; VI-NEXT:    s_and_b32 s29, s29, 0xffff
; VI-NEXT:    s_and_b32 s28, s28, 0xffff
; VI-NEXT:    s_and_b32 s27, s27, 0xffff
; VI-NEXT:    s_and_b32 s26, s26, 0xffff
; VI-NEXT:    s_and_b32 s25, s25, 0xffff
; VI-NEXT:    s_and_b32 s24, s24, 0xffff
; VI-NEXT:    s_and_b32 s23, s23, 0xffff
; VI-NEXT:    s_and_b32 s22, s22, 0xffff
; VI-NEXT:    s_and_b32 s21, s21, 0xffff
; VI-NEXT:    s_and_b32 s20, s20, 0xffff
; VI-NEXT:    s_and_b32 s19, s19, 0xffff
; VI-NEXT:    s_and_b32 s18, s18, 0xffff
; VI-NEXT:    s_and_b32 s17, s17, 0xffff
; VI-NEXT:    s_and_b32 s16, s16, 0xffff
; VI-NEXT:    s_and_b32 s8, s8, 0xffff
; VI-NEXT:    s_and_b32 s9, s9, 0xffff
; VI-NEXT:    s_and_b32 s10, s10, 0xffff
; VI-NEXT:    s_and_b32 s11, s11, 0xffff
; VI-NEXT:    s_and_b32 s12, s12, 0xffff
; VI-NEXT:    s_and_b32 s13, s13, 0xffff
; VI-NEXT:    s_and_b32 s14, s14, 0xffff
; VI-NEXT:    s_and_b32 s15, s15, 0xffff
; VI-NEXT:    s_and_b32 s40, s40, 0xffff
; VI-NEXT:    s_and_b32 s41, s41, 0xffff
; VI-NEXT:    s_and_b32 s42, s42, 0xffff
; VI-NEXT:    s_and_b32 s43, s43, 0xffff
; VI-NEXT:    s_and_b32 s44, s44, 0xffff
; VI-NEXT:    s_and_b32 s45, s45, 0xffff
; VI-NEXT:    s_and_b32 s46, s46, 0xffff
; VI-NEXT:    s_and_b32 s5, s5, 0xffff
; VI-NEXT:    s_or_b32 s6, s39, s6
; VI-NEXT:    s_or_b32 s7, s38, s7
; VI-NEXT:    s_or_b32 s29, s37, s29
; VI-NEXT:    s_or_b32 s28, s36, s28
; VI-NEXT:    s_or_b32 s27, s35, s27
; VI-NEXT:    s_or_b32 s26, s34, s26
; VI-NEXT:    s_or_b32 s25, s31, s25
; VI-NEXT:    s_or_b32 s24, s30, s24
; VI-NEXT:    s_or_b32 s23, vcc_hi, s23
; VI-NEXT:    s_or_b32 s22, vcc_lo, s22
; VI-NEXT:    s_or_b32 s21, s91, s21
; VI-NEXT:    s_or_b32 s20, s90, s20
; VI-NEXT:    s_or_b32 s19, s89, s19
; VI-NEXT:    s_or_b32 s18, s88, s18
; VI-NEXT:    s_or_b32 s17, s79, s17
; VI-NEXT:    s_or_b32 s16, s78, s16
; VI-NEXT:    s_or_b32 s8, s77, s8
; VI-NEXT:    s_or_b32 s9, s76, s9
; VI-NEXT:    s_or_b32 s10, s75, s10
; VI-NEXT:    s_or_b32 s11, s74, s11
; VI-NEXT:    s_or_b32 s12, s73, s12
; VI-NEXT:    s_or_b32 s13, s72, s13
; VI-NEXT:    s_or_b32 s14, s63, s14
; VI-NEXT:    s_or_b32 s15, s62, s15
; VI-NEXT:    s_or_b32 s40, s61, s40
; VI-NEXT:    s_or_b32 s41, s60, s41
; VI-NEXT:    s_or_b32 s42, s59, s42
; VI-NEXT:    s_or_b32 s43, s58, s43
; VI-NEXT:    s_or_b32 s44, s57, s44
; VI-NEXT:    s_or_b32 s45, s56, s45
; VI-NEXT:    s_or_b32 s46, s47, s46
; VI-NEXT:    s_or_b32 s4, s4, s5
; VI-NEXT:    s_add_i32 s6, s6, 0x30000
; VI-NEXT:    s_add_i32 s7, s7, 0x30000
; VI-NEXT:    s_add_i32 s29, s29, 0x30000
; VI-NEXT:    s_add_i32 s28, s28, 0x30000
; VI-NEXT:    s_add_i32 s27, s27, 0x30000
; VI-NEXT:    s_add_i32 s26, s26, 0x30000
; VI-NEXT:    s_add_i32 s25, s25, 0x30000
; VI-NEXT:    s_add_i32 s24, s24, 0x30000
; VI-NEXT:    s_add_i32 s23, s23, 0x30000
; VI-NEXT:    s_add_i32 s22, s22, 0x30000
; VI-NEXT:    s_add_i32 s21, s21, 0x30000
; VI-NEXT:    s_add_i32 s20, s20, 0x30000
; VI-NEXT:    s_add_i32 s19, s19, 0x30000
; VI-NEXT:    s_add_i32 s18, s18, 0x30000
; VI-NEXT:    s_add_i32 s17, s17, 0x30000
; VI-NEXT:    s_add_i32 s16, s16, 0x30000
; VI-NEXT:    s_add_i32 s8, s8, 0x30000
; VI-NEXT:    s_add_i32 s9, s9, 0x30000
; VI-NEXT:    s_add_i32 s10, s10, 0x30000
; VI-NEXT:    s_add_i32 s11, s11, 0x30000
; VI-NEXT:    s_add_i32 s12, s12, 0x30000
; VI-NEXT:    s_add_i32 s13, s13, 0x30000
; VI-NEXT:    s_add_i32 s14, s14, 0x30000
; VI-NEXT:    s_add_i32 s15, s15, 0x30000
; VI-NEXT:    s_add_i32 s40, s40, 0x30000
; VI-NEXT:    s_add_i32 s41, s41, 0x30000
; VI-NEXT:    s_add_i32 s42, s42, 0x30000
; VI-NEXT:    s_add_i32 s43, s43, 0x30000
; VI-NEXT:    s_add_i32 s44, s44, 0x30000
; VI-NEXT:    s_add_i32 s45, s45, 0x30000
; VI-NEXT:    s_add_i32 s46, s46, 0x30000
; VI-NEXT:    s_add_i32 s47, s4, 0x30000
; VI-NEXT:  .LBB111_3: ; %end
; VI-NEXT:    v_mov_b32_e32 v0, s16
; VI-NEXT:    v_mov_b32_e32 v1, s17
; VI-NEXT:    v_mov_b32_e32 v2, s18
; VI-NEXT:    v_mov_b32_e32 v3, s19
; VI-NEXT:    v_mov_b32_e32 v4, s20
; VI-NEXT:    v_mov_b32_e32 v5, s21
; VI-NEXT:    v_mov_b32_e32 v6, s22
; VI-NEXT:    v_mov_b32_e32 v7, s23
; VI-NEXT:    v_mov_b32_e32 v8, s24
; VI-NEXT:    v_mov_b32_e32 v9, s25
; VI-NEXT:    v_mov_b32_e32 v10, s26
; VI-NEXT:    v_mov_b32_e32 v11, s27
; VI-NEXT:    v_mov_b32_e32 v12, s28
; VI-NEXT:    v_mov_b32_e32 v13, s29
; VI-NEXT:    v_mov_b32_e32 v14, s7
; VI-NEXT:    v_mov_b32_e32 v15, s6
; VI-NEXT:    v_mov_b32_e32 v16, s47
; VI-NEXT:    v_mov_b32_e32 v17, s46
; VI-NEXT:    v_mov_b32_e32 v18, s45
; VI-NEXT:    v_mov_b32_e32 v19, s44
; VI-NEXT:    v_mov_b32_e32 v20, s43
; VI-NEXT:    v_mov_b32_e32 v21, s42
; VI-NEXT:    v_mov_b32_e32 v22, s41
; VI-NEXT:    v_mov_b32_e32 v23, s40
; VI-NEXT:    v_mov_b32_e32 v24, s15
; VI-NEXT:    v_mov_b32_e32 v25, s14
; VI-NEXT:    v_mov_b32_e32 v26, s13
; VI-NEXT:    v_mov_b32_e32 v27, s12
; VI-NEXT:    v_mov_b32_e32 v28, s11
; VI-NEXT:    v_mov_b32_e32 v29, s10
; VI-NEXT:    v_mov_b32_e32 v30, s9
; VI-NEXT:    v_mov_b32_e32 v31, s8
; VI-NEXT:    v_readlane_b32 s39, v32, 7
; VI-NEXT:    v_readlane_b32 s38, v32, 6
; VI-NEXT:    v_readlane_b32 s37, v32, 5
; VI-NEXT:    v_readlane_b32 s36, v32, 4
; VI-NEXT:    v_readlane_b32 s35, v32, 3
; VI-NEXT:    v_readlane_b32 s34, v32, 2
; VI-NEXT:    v_readlane_b32 s31, v32, 1
; VI-NEXT:    v_readlane_b32 s30, v32, 0
; VI-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; VI-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; VI-NEXT:    s_mov_b64 exec, s[4:5]
; VI-NEXT:    s_waitcnt vmcnt(0)
; VI-NEXT:    s_setpc_b64 s[30:31]
; VI-NEXT:  .LBB111_4:
; VI-NEXT:    s_branch .LBB111_2
;
; GFX9-LABEL: bitcast_v64i16_to_v64f16_scalar:
; GFX9:       ; %bb.0:
; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_store_dword v32, off, s[0:3], s32 ; 4-byte Folded Spill
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    v_writelane_b32 v32, s30, 0
; GFX9-NEXT:    v_writelane_b32 v32, s31, 1
; GFX9-NEXT:    v_writelane_b32 v32, s36, 2
; GFX9-NEXT:    v_writelane_b32 v32, s37, 3
; GFX9-NEXT:    v_writelane_b32 v32, s38, 4
; GFX9-NEXT:    v_writelane_b32 v32, s39, 5
; GFX9-NEXT:    v_writelane_b32 v32, s48, 6
; GFX9-NEXT:    v_writelane_b32 v32, s49, 7
; GFX9-NEXT:    v_writelane_b32 v32, s50, 8
; GFX9-NEXT:    v_readfirstlane_b32 s4, v18
; GFX9-NEXT:    v_writelane_b32 v32, s51, 9
; GFX9-NEXT:    v_readfirstlane_b32 s51, v17
; GFX9-NEXT:    v_readfirstlane_b32 s50, v16
; GFX9-NEXT:    v_readfirstlane_b32 s49, v15
; GFX9-NEXT:    v_readfirstlane_b32 s48, v14
; GFX9-NEXT:    v_readfirstlane_b32 s47, v13
; GFX9-NEXT:    v_readfirstlane_b32 s46, v12
; GFX9-NEXT:    v_readfirstlane_b32 s45, v11
; GFX9-NEXT:    v_readfirstlane_b32 s44, v10
; GFX9-NEXT:    v_readfirstlane_b32 s43, v9
; GFX9-NEXT:    v_readfirstlane_b32 s42, v8
; GFX9-NEXT:    v_readfirstlane_b32 s41, v7
; GFX9-NEXT:    v_readfirstlane_b32 s40, v6
; GFX9-NEXT:    v_readfirstlane_b32 s39, v5
; GFX9-NEXT:    v_readfirstlane_b32 s38, v4
; GFX9-NEXT:    v_readfirstlane_b32 s37, v3
; GFX9-NEXT:    v_readfirstlane_b32 s36, v2
; GFX9-NEXT:    v_readfirstlane_b32 s31, v1
; GFX9-NEXT:    s_cmp_lg_u32 s4, 0
; GFX9-NEXT:    v_readfirstlane_b32 s30, v0
; GFX9-NEXT:    s_cbranch_scc0 .LBB111_3
; GFX9-NEXT:  ; %bb.1: ; %cmp.false
; GFX9-NEXT:    s_cbranch_execnz .LBB111_4
; GFX9-NEXT:  .LBB111_2: ; %cmp.true
; GFX9-NEXT:    v_pk_add_u16 v15, s31, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v14, s30, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v13, s29, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v12, s28, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v11, s27, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v10, s26, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v9, s25, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v8, s24, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v7, s23, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v6, s22, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v5, s21, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v4, s20, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v3, s19, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v2, s18, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v1, s17, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v0, s16, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v31, s51, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v30, s50, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v29, s49, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v28, s48, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v27, s47, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v26, s46, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v25, s45, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v24, s44, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v23, s43, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v22, s42, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v21, s41, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v20, s40, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v19, s39, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v18, s38, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v17, s37, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    v_pk_add_u16 v16, s36, 3 op_sel_hi:[1,0]
; GFX9-NEXT:    s_branch .LBB111_5
; GFX9-NEXT:  .LBB111_3:
; GFX9-NEXT:    s_branch .LBB111_2
; GFX9-NEXT:  .LBB111_4:
; GFX9-NEXT:    v_mov_b32_e32 v16, s36
; GFX9-NEXT:    v_mov_b32_e32 v0, s16
; GFX9-NEXT:    v_mov_b32_e32 v17, s37
; GFX9-NEXT:    v_mov_b32_e32 v18, s38
; GFX9-NEXT:    v_mov_b32_e32 v19, s39
; GFX9-NEXT:    v_mov_b32_e32 v20, s40
; GFX9-NEXT:    v_mov_b32_e32 v21, s41
; GFX9-NEXT:    v_mov_b32_e32 v22, s42
; GFX9-NEXT:    v_mov_b32_e32 v23, s43
; GFX9-NEXT:    v_mov_b32_e32 v24, s44
; GFX9-NEXT:    v_mov_b32_e32 v25, s45
; GFX9-NEXT:    v_mov_b32_e32 v26, s46
; GFX9-NEXT:    v_mov_b32_e32 v27, s47
; GFX9-NEXT:    v_mov_b32_e32 v28, s48
; GFX9-NEXT:    v_mov_b32_e32 v29, s49
; GFX9-NEXT:    v_mov_b32_e32 v30, s50
; GFX9-NEXT:    v_mov_b32_e32 v31, s51
; GFX9-NEXT:    v_mov_b32_e32 v1, s17
; GFX9-NEXT:    v_mov_b32_e32 v2, s18
; GFX9-NEXT:    v_mov_b32_e32 v3, s19
; GFX9-NEXT:    v_mov_b32_e32 v4, s20
; GFX9-NEXT:    v_mov_b32_e32 v5, s21
; GFX9-NEXT:    v_mov_b32_e32 v6, s22
; GFX9-NEXT:    v_mov_b32_e32 v7, s23
; GFX9-NEXT:    v_mov_b32_e32 v8, s24
; GFX9-NEXT:    v_mov_b32_e32 v9, s25
; GFX9-NEXT:    v_mov_b32_e32 v10, s26
; GFX9-NEXT:    v_mov_b32_e32 v11, s27
; GFX9-NEXT:    v_mov_b32_e32 v12, s28
; GFX9-NEXT:    v_mov_b32_e32 v13, s29
; GFX9-NEXT:    v_mov_b32_e32 v14, s30
; GFX9-NEXT:    v_mov_b32_e32 v15, s31
; GFX9-NEXT:  .LBB111_5: ; %end
; GFX9-NEXT:    v_readlane_b32 s51, v32, 9
; GFX9-NEXT:    v_readlane_b32 s50, v32, 8
; GFX9-NEXT:    v_readlane_b32 s49, v32, 7
; GFX9-NEXT:    v_readlane_b32 s48, v32, 6
; GFX9-NEXT:    v_readlane_b32 s39, v32, 5
; GFX9-NEXT:    v_readlane_b32 s38, v32, 4
; GFX9-NEXT:    v_readlane_b32 s37, v32, 3
; GFX9-NEXT:    v_readlane_b32 s36, v32, 2
; GFX9-NEXT:    v_readlane_b32 s31, v32, 1
; GFX9-NEXT:    v_readlane_b32 s30, v32, 0
; GFX9-NEXT:    s_xor_saveexec_b64 s[4:5], -1
; GFX9-NEXT:    buffer_load_dword v32, off, s[0:3], s32 ; 4-byte Folded Reload
; GFX9-NEXT:    s_mov_b64 exec, s[4:5]
; GFX9-NEXT:    s_waitcnt vmcnt(0)
; GFX9-NEXT:    s_setpc_b64 s[30:31]
;
; GFX11-LABEL: bitcast_v64i16_to_v64f16_scalar:
; GFX11:       ; %bb.0:
; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-NEXT:    s_xor_saveexec_b32 s4, -1
; GFX11-NEXT:    scratch_store_b32 off, v32, s32 ; 4-byte Folded Spill
; GFX11-NEXT:    s_mov_b32 exec_lo, s4
; GFX11-NEXT:    v_writelane_b32 v32, s36, 0
; GFX11-NEXT:    s_mov_b32 s12, s0
; GFX11-NEXT:    v_readfirstlane_b32 s0, v14
; GFX11-NEXT:    v_readfirstlane_b32 s47, v9
; GFX11-NEXT:    v_readfirstlane_b32 s46, v8
; GFX11-NEXT:    v_writelane_b32 v32, s37, 1
; GFX11-NEXT:    v_readfirstlane_b32 s45, v7
; GFX11-NEXT:    v_readfirstlane_b32 s44, v6
; GFX11-NEXT:    v_readfirstlane_b32 s43, v5
; GFX11-NEXT:    v_readfirstlane_b32 s42, v4
; GFX11-NEXT:    v_writelane_b32 v32, s38, 2
; GFX11-NEXT:    v_readfirstlane_b32 s41, v3
; GFX11-NEXT:    v_readfirstlane_b32 s40, v2
; GFX11-NEXT:    v_readfirstlane_b32 s38, v0
; GFX11-NEXT:    s_mov_b32 s37, s29
; GFX11-NEXT:    v_writelane_b32 v32, s39, 3
; GFX11-NEXT:    v_readfirstlane_b32 s39, v1
; GFX11-NEXT:    s_mov_b32 s36, s28
; GFX11-NEXT:    s_mov_b32 s15, s3
; GFX11-NEXT:    s_mov_b32 s14, s2
; GFX11-NEXT:    v_writelane_b32 v32, s48, 4
; GFX11-NEXT:    v_readfirstlane_b32 s48, v10
; GFX11-NEXT:    s_mov_b32 s13, s1
; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
; GFX11-NEXT:    s_mov_b32 s0, 0
; GFX11-NEXT:    v_writelane_b32 v32, s49, 5
; GFX11-NEXT:    v_readfirstlane_b32 s49, v11
; GFX11-NEXT:    v_writelane_b32 v32, s50, 6
; GFX11-NEXT:    v_readfirstlane_b32 s50, v12
; GFX11-NEXT:    v_writelane_b32 v32, s51, 7
; GFX11-NEXT:    v_readfirstlane_b32 s51, v13
; GFX11-NEXT:    s_cbranch_scc0 .LBB111_3
; GFX11-NEXT:  ; %bb.1: ; %Flow
; GFX11-NEXT:    s_and_not1_b32 vcc_lo, exec_lo, s0
; GFX11-NEXT:    s_cbranch_vccnz .LBB111_4
; GFX11-NEXT:  .LBB111_2: ; %cmp.true
; GFX11-NEXT:    v_pk_add_u16 v15, s27, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v14, s26, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v13, s25, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v12, s24, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v11, s23, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v10, s22, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v9, s21, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v8, s20, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v7, s19, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v6, s18, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v5, s17, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v4, s16, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v3, s15, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v2, s14, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v1, s13, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v0, s12, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v31, s51, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v30, s50, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v29, s49, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v28, s48, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v27, s47, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v26, s46, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v25, s45, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v24, s44, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v23, s43, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v22, s42, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v21, s41, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v20, s40, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v19, s39, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v18, s38, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v17, s37, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    v_pk_add_u16 v16, s36, 3 op_sel_hi:[1,0]
; GFX11-NEXT:    s_branch .LBB111_5
; GFX11-NEXT:  .LBB111_3:
; GFX11-NEXT:    s_branch .LBB111_2
; GFX11-NEXT:  .LBB111_4:
; GFX11-NEXT:    v_dual_mov_b32 v16, s36 :: v_dual_mov_b32 v17, s37
; GFX11-NEXT:    v_dual_mov_b32 v0, s12 :: v_dual_mov_b32 v1, s13
; GFX11-NEXT:    v_dual_mov_b32 v18, s38 :: v_dual_mov_b32 v19, s39
; GFX11-NEXT:    v_dual_mov_b32 v20, s40 :: v_dual_mov_b32 v21, s41
; GFX11-NEXT:    v_dual_mov_b32 v22, s42 :: v_dual_mov_b32 v23, s43
; GFX11-NEXT:    v_dual_mov_b32 v24, s44 :: v_dual_mov_b32 v25, s45
; GFX11-NEXT:    v_dual_mov_b32 v26, s46 :: v_dual_mov_b32 v27, s47
; GFX11-NEXT:    v_dual_mov_b32 v28, s48 :: v_dual_mov_b32 v29, s49
; GFX11-NEXT:    v_dual_mov_b32 v30, s50 :: v_dual_mov_b32 v31, s51
; GFX11-NEXT:    v_dual_mov_b32 v2, s14 :: v_dual_mov_b32 v3, s15
; GFX11-NEXT:    v_dual_mov_b32 v4, s16 :: v_dual_mov_b32 v5, s17
; GFX11-NEXT:    v_dual_mov_b32 v6, s18 :: v_dual_mov_b32 v7, s19
; GFX11-NEXT:    v_dual_mov_b32 v8, s20 :: v_dual_mov_b32 v9, s21
; GFX11-NEXT:    v_dual_mov_b32 v10, s22 :: v_dual_mov_b32 v11, s23
; GFX11-NEXT:    v_dual_mov_b32 v12, s24 :: v_dual_mov_b32 v13, s25
; GFX11-NEXT:    v_dual_mov_b32 v14, s26 :: v_dual_mov_b32 v15, s27
; GFX11-NEXT:  .LBB111_5: ; %end
; GFX11-NEXT:    v_readlane_b32 s51, v32, 7
; GFX11-NEXT:    v_readlane_b32 s50, v32, 6
; GFX11-NEXT:    v_readlane_b32 s49, v32, 5
; GFX11-NEXT:    v_readlane_b32 s48, v32, 4
; GFX11-NEXT:    v_readlane_b32 s39, v32, 3
; GFX11-NEXT:    v_readlane_b32 s38, v32, 2
; GFX11-NEXT:    v_readlane_b32 s37, v32, 1
; GFX11-NEXT:    v_readlane_b32 s36, v32, 0
; GFX11-NEXT:    s_xor_saveexec_b32 s0, -1
; GFX11-NEXT:    scratch_load_b32 v32, off, s32 ; 4-byte Folded Reload
; GFX11-NEXT:    s_mov_b32 exec_lo, s0
; GFX11-NEXT:    s_waitcnt vmcnt(0)
; GFX11-NEXT:    s_setpc_b64 s[30:31]
  %cmp = icmp eq i32 %b, 0
  br i1 %cmp, label %cmp.true, label %cmp.false

cmp.true:
  %a1 = add <64 x i16> %a, splat (i16 3)
  %a2 = bitcast <64 x i16> %a1 to <64 x half>
  br label %end

cmp.false:
  %a3 = bitcast <64 x i16> %a to <64 x half>
  br label %end

end:
  %phi = phi <64 x half> [ %a2, %cmp.true ], [ %a3, %cmp.false ]
  ret <64 x half> %phi
}